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

23
vllm/model_executor/models/qwen3_next_mtp.py
View File

@@ -15,7 +15,6 @@ from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.linear import ColumnParallelLinear
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.vocab_parallel_embedding import (
DEFAULT_VOCAB_PADDING_SIZE,
ParallelLMHead,
VocabParallelEmbedding,
)
@@ -48,17 +47,12 @@ class Qwen3NextMultiTokenPredictor(nn.Module):
model_config = vllm_config.model_config
quant_config = vllm_config.quant_config
lora_config = vllm_config.lora_config
config: Qwen3NextConfig = model_config.hf_config
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
self.mtp_start_layer_idx = config.num_hidden_layers
self.num_mtp_layers = getattr(config, "num_nextn_predict_layers", 1)
@@ -66,7 +60,6 @@ class Qwen3NextMultiTokenPredictor(nn.Module):
self.embed_tokens = VocabParallelEmbedding(
self.vocab_size,
config.hidden_size,
org_num_embeddings=config.vocab_size,
)
self.fc = ColumnParallelLinear(
@@ -252,17 +245,13 @@ class Qwen3NextMTP(nn.Module, SupportsPP, QwenNextMixtureOfExperts):
self.model = Qwen3NextMultiTokenPredictor(
vllm_config=vllm_config, prefix=maybe_prefix(prefix, "mtp")
)
self.unpadded_vocab_size = config.vocab_size
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,
prefix=maybe_prefix(prefix, "lm_head"),
)
self.logits_processor = LogitsProcessor(
self.unpadded_vocab_size, config.vocab_size
)
self.logits_processor = LogitsProcessor(config.vocab_size)
self.make_empty_intermediate_tensors = (
self.model.make_empty_intermediate_tensors
)