biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[LoRA] Support dynamically initialize packed_modules_mapping for VLM with arbitrary components (#18987)

Browse Source 
Signed-off-by: isotr0py <2037008807@qq.com>
Signed-off-by: Isotr0py <2037008807@qq.com>
This commit is contained in:
Isotr0py
2025-06-01 11:06:57 +08:00
committed by GitHub
parent 6aa8f9a4e7
commit a35ca765a5
7 changed files with 32 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/model_executor/models/intern_vit.py
View File

@@ -415,6 +415,10 @@ class InternVisionEncoder(nn.Module):
class InternVisionModel(nn.Module):
packed_modules_mapping = {
"qkv": ["qkv"],
}
def __init__(
self,
config: PretrainedConfig,

9
vllm/model_executor/models/internvl.py
View File

@@ -1019,15 +1019,6 @@ class InternVLMultiModalProcessor(
class InternVLChatModel(nn.Module, SupportsMultiModal, SupportsPP,
SupportsLoRA):
packed_modules_mapping = {
"wqkv": ["wqkv"],
"qkv": ["qkv"],
"gate_up_proj": [
"w1",
"w3",
],
}
def __init__(self, *, vllm_config: VllmConfig, prefix: str = "") -> None:
super().__init__()

11
vllm/model_executor/models/qwen2_5_vl.py
View File

@@ -821,17 +821,6 @@ class Qwen2_5_VLMultiModalProcessor(Qwen2VLMultiModalProcessor):
dummy_inputs=Qwen2_5_VLDummyInputsBuilder)
class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
SupportsLoRA, SupportsPP):
packed_modules_mapping = {
"qkv_proj": [
"q_proj",
"k_proj",
"v_proj",
],
"gate_up_proj": [
"gate_proj",
"up_proj",
],
}
# To ensure correct weight loading and mapping.
hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={

11
vllm/model_executor/models/qwen2_vl.py
View File

@@ -1069,17 +1069,6 @@ class Qwen2VLMultiModalProcessor(BaseMultiModalProcessor[Qwen2VLProcessingInfo]
dummy_inputs=Qwen2VLDummyInputsBuilder)
class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
SupportsLoRA, SupportsPP):
packed_modules_mapping = {
"qkv_proj": [
"q_proj",
"k_proj",
"v_proj",
],
"gate_up_proj": [
"gate_proj",
"up_proj",
],
}
# To ensure correct weight loading and mapping.
hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={