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

28
vllm/model_executor/models/zamba2.py
View File

@@ -38,7 +38,6 @@ from vllm.model_executor.layers.mamba.mamba_utils import (
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,
)
@@ -692,19 +691,13 @@ class Zamba2Model(nn.Module):
assert not is_lora_enabled
self.config = 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
# Initialize token embeddings
self.embed_tokens = VocabParallelEmbedding(
self.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
)
# Map hybrid layer indices to block indices
@@ -911,7 +904,7 @@ class Zamba2ForCausalLM(nn.Module, HasInnerState, IsHybrid, SupportsMambaPrefixC
(not supported by Mamba)
"""
config = vllm_config.model_config.hf_config
lora_config = vllm_config.lora_config
scheduler_config = vllm_config.scheduler_config
super().__init__()
@@ -919,9 +912,6 @@ class Zamba2ForCausalLM(nn.Module, HasInnerState, IsHybrid, SupportsMambaPrefixC
self.vllm_config = vllm_config
self.scheduler_config = scheduler_config
self.model_config = vllm_config.model_config
self.unpadded_vocab_size = config.vocab_size
if lora_config:
self.unpadded_vocab_size += lora_config.lora_extra_vocab_size
# Initialize core model
self.model = Zamba2Model(
@@ -930,23 +920,15 @@ class Zamba2ForCausalLM(nn.Module, HasInnerState, IsHybrid, SupportsMambaPrefixC
# Initialize language modeling head
self.lm_head = ParallelLMHead(
self.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,
prefix=maybe_prefix(prefix, "lm_head"),
)
# Tie weights with input embeddings if using same dimensions
self.lm_head = self.lm_head.tie_weights(self.model.embed_tokens)
# Initialize logits processing and sampling
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, config.vocab_size
)
self.logits_processor = LogitsProcessor(config.vocab_size)
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
"""Convert input token IDs to embeddings.