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

30
vllm/model_executor/models/minicpm.py
View File

@@ -55,7 +55,6 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.rotary_embedding import get_rope
from vllm.model_executor.layers.vocab_parallel_embedding import (
DEFAULT_VOCAB_PADDING_SIZE,
ParallelLMHead,
VocabParallelEmbedding,
)
@@ -405,22 +404,16 @@ class MiniCPMModel(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
self.cache_config = cache_config
self.quant_config = quant_config
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.org_vocab_size = config.vocab_size
self.vocab_size = config.vocab_size
self.embed_tokens = VocabParallelEmbedding(
self.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
)
self.num_experts = getattr(self.config, "num_experts", 0)
self._init_layers(prefix, config, cache_config, quant_config)
@@ -588,13 +581,13 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
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
parallel_config = vllm_config.parallel_config
self.prefix = prefix
self.vllm_config = vllm_config
self.config = config
self.lora_config = lora_config
self.cache_config = cache_config
self.quant_config = quant_config
@@ -602,18 +595,9 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
)
unpadded_vocab_size = config.vocab_size
if lora_config:
unpadded_vocab_size += lora_config.lora_extra_vocab_size
self.lm_head = ParallelLMHead(
unpadded_vocab_size,
config.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,
prefix=maybe_prefix(prefix, "lm_head"),
)
@@ -621,7 +605,7 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
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, config.vocab_size)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors
)