[Model][LoRA] Enable LoRA support for VoyageQwen3BidirectionalEmbedModel (#57708)
Signed-off-by: Mahir Dursunoglu <142054011+mahird3@users.noreply.github.com>
mahird3 committed
Sep 20, 2026 at 11:50 UTC
0748d3bd57cb5c4303e58c8465e21f069629480d
1 file changed
+7
-1
vllm/model_executor/models/voyage.py
+7
-1
@@ -8,6 +8,7 @@ import torch
8
import torch.nn as nn
9
10
from vllm.config import VllmConfig
11
+from vllm.model_executor.models.interfaces import SupportsLoRA
12
from vllm.model_executor.models.qwen3 import Qwen3Model
13
from vllm.model_executor.models.utils import AutoWeightsLoader, maybe_prefix
14
from vllm.sequence import IntermediateTensors
@@ -15,7 +16,7 @@ from vllm.sequence import IntermediateTensors
16
WeightItem = tuple[str, torch.Tensor]
17
18
18
-class VoyageQwen3BidirectionalEmbedModel(nn.Module):
19
+class VoyageQwen3BidirectionalEmbedModel(nn.Module, SupportsLoRA):
20
"""Qwen3Model + Voyage embedding head + bidirectional attention.
21
22
Checkpoint conventions (HF):
@@ -30,6 +31,11 @@ class VoyageQwen3BidirectionalEmbedModel(nn.Module):
31
- No "model." prefix
32
"""
33
34
+ packed_modules_mapping = {
35
+ "qkv_proj": ["q_proj", "k_proj", "v_proj"],
36
+ "gate_up_proj": ["gate_proj", "up_proj"],
37
+ }
38
+
39
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
40
super().__init__()
41
self.config = vllm_config.model_config.hf_config