[Model][Quant] Fix GLM, Fix fused module mappings for quantization (#12634)
Signed-off-by: mgoin <michael@neuralmagic.com> Signed-off-by: Kyle Sayers <kylesayrs@gmail.com> Co-authored-by: mgoin <michael@neuralmagic.com>
This commit is contained in:
@@ -1473,6 +1473,7 @@ class MiniCPMV(MiniCPMVBaseModel, SupportsMultiModal, SupportsLoRA):
|
||||
"""
|
||||
# Ensure that the LoRA support check passes when the class is not
|
||||
# initialized, but set all these attributes to empty.
|
||||
# These will be updated when an instance class is selected
|
||||
packed_modules_mapping = {}
|
||||
supported_lora_modules = []
|
||||
embedding_modules = {}
|
||||
@@ -1489,8 +1490,15 @@ class MiniCPMV(MiniCPMVBaseModel, SupportsMultiModal, SupportsLoRA):
|
||||
version = str(config.version).split(".")
|
||||
version = tuple([int(x) for x in version])
|
||||
# Dispatch class based on version
|
||||
instance_class = _SUPPORT_VERSION.get(version)
|
||||
if instance_class is None:
|
||||
instance_cls = _SUPPORT_VERSION.get(version)
|
||||
if instance_cls is None:
|
||||
raise ValueError(
|
||||
"Currently, MiniCPMV only supports versions 2.0, 2.5, and 2.6")
|
||||
return instance_class(vllm_config=vllm_config, prefix=prefix)
|
||||
|
||||
# quant_config references base class members,
|
||||
# so update values before init is called
|
||||
cls.packed_modules_mapping.update(instance_cls.packed_modules_mapping)
|
||||
cls.supported_lora_modules += instance_cls.supported_lora_modules
|
||||
cls.embedding_modules.update(instance_cls.embedding_modules)
|
||||
cls.embedding_padding_modules += instance_cls.embedding_padding_modules
|
||||
return instance_cls(vllm_config=vllm_config, prefix=prefix)
|
||||
|
||||
Reference in New Issue
Block a user