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

[Bugfix]Fix MiniCPM's LoRA bug (#9286)

Browse Source
This commit is contained in:
Jee Jee Li
2024-10-13 00:36:47 +08:00
committed by GitHub
parent 2b184ddd4f
commit 250e26a63e
3 changed files with 40 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
vllm/model_executor/models/minicpm.py
View File

@@ -474,17 +474,18 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
unpadded_vocab_size = config.vocab_size
if lora_config:
unpadded_vocab_size += lora_config.lora_extra_vocab_size
if not self.config.tie_word_embeddings:
self.lm_head = ParallelLMHead(
unpadded_vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
padding_size=DEFAULT_VOCAB_PADDING_SIZE
# We need bigger padding if using lora for kernel
# compatibility
if not lora_config else lora_config.lora_vocab_padding_size,
quant_config=quant_config,
)
self.lm_head = ParallelLMHead(
unpadded_vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
padding_size=DEFAULT_VOCAB_PADDING_SIZE
# We need bigger padding if using lora for kernel
# compatibility
if not lora_config else lora_config.lora_vocab_padding_size,
quant_config=quant_config,
)
if config.tie_word_embeddings:
self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
self.scale_width = self.config.hidden_size / self.config.dim_model_base
self.logits_processor = LogitsProcessor(unpadded_vocab_size,
@@ -517,11 +518,7 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
sampling_metadata: SamplingMetadata,
) -> Optional[torch.Tensor]:
hidden_states = hidden_states / self.scale_width
if self.config.tie_word_embeddings:
lm_head = self.model.embed_tokens
else:
lm_head = self.lm_head
logits = self.logits_processor(lm_head, hidden_states,
logits = self.logits_processor(self.lm_head, hidden_states,
sampling_metadata)
return logits