[Model][LoRA] Enable LoRA support for BERT and RoBERTa embedding models (#61033)

Signed-off-by: hoaivoke <vokehoai@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Hoai Vo
2026-10-12 14:09:53 +00:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 716961cf1f
commit f3cb10b690
2 changed files with 18 additions and 5 deletions
+15 -2
View File
@@ -41,7 +41,7 @@ from vllm.sequence import IntermediateTensors
from vllm.tasks import PoolingTask
from vllm.v1.pool.metadata import PoolingMetadata
from .interfaces import SupportsCrossEncoding, SupportsQuant
from .interfaces import SupportsCrossEncoding, SupportsLoRA, SupportsQuant
from .interfaces_base import attn_type, default_pooling_type
from .utils import AutoWeightsLoader, WeightsMapper, maybe_prefix
@@ -448,7 +448,7 @@ class BertPoolingModel(BertModel):
@default_pooling_type(seq_pooling_type="CLS")
class BertEmbeddingModel(nn.Module, SupportsQuant):
class BertEmbeddingModel(nn.Module, SupportsQuant, SupportsLoRA):
"""A model that uses Bert to provide embedding functionalities.
This class encapsulates the BertModel and provides an interface for
@@ -463,9 +463,22 @@ class BertEmbeddingModel(nn.Module, SupportsQuant):
is_pooling_model = True
embedding_class: type[nn.Module] = BertEmbedding
packed_modules_mapping = {"qkv_proj": ["query", "key", "value"]}
# PEFT adapters name modules relative to the HF backbone (optionally under
# `bert.`/`roberta.`), while vLLM nests the backbone under `model.`.
hf_to_vllm_mapper = WeightsMapper(
orig_to_new_prefix={
"bert.": "model.",
"roberta.": "model.",
"encoder.": "model.encoder.",
"embeddings.": "model.embeddings.",
}
)
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
self.config = vllm_config.model_config.hf_config
pooler_config = vllm_config.model_config.pooler_config
assert pooler_config is not None