[V1]SupportsV0Only protocol for model definitions (#13959)
Signed-off-by: Roger Wang <ywang@roblox.com>
This commit is contained in:
@@ -63,7 +63,7 @@ from vllm.multimodal.processing import (BaseProcessingInfo,
|
||||
from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
|
||||
|
||||
from .clip import CLIPMLP
|
||||
from .interfaces import SupportsMultiModal
|
||||
from .interfaces import SupportsMultiModal, SupportsV0Only
|
||||
from .llama import LlamaDecoderLayer, LlamaMLP
|
||||
from .utils import maybe_prefix
|
||||
|
||||
@@ -1128,7 +1128,8 @@ class MllamaForCausalLM(nn.Module):
|
||||
@MULTIMODAL_REGISTRY.register_processor(MllamaMultiModalProcessor,
|
||||
info=MllamaProcessingInfo,
|
||||
dummy_inputs=MllamaDummyInputsBuilder)
|
||||
class MllamaForConditionalGeneration(nn.Module, SupportsMultiModal):
|
||||
class MllamaForConditionalGeneration(nn.Module, SupportsMultiModal,
|
||||
SupportsV0Only):
|
||||
packed_modules_mapping = {
|
||||
"qkv_proj": ["q_proj", "k_proj", "v_proj"],
|
||||
"gate_up_proj": ["gate_proj", "up_proj"]
|
||||
|
||||
Reference in New Issue
Block a user