rixty/KoboldAI-Client
1
1
Fork 0
mirror of https://github.com/KoboldAI/KoboldAI-Client.git synced 2025-06-05 21:59:24 +02:00
Code Issues Projects Releases Wiki Activity

Restore choice of lazyload or not

Browse Source
This commit is contained in:
somebody
2023-06-21 14:13:14 -05:00
parent 0052ad401a
commit 5f224e1366
1 changed files with 9 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
modeling/inference_models/hf_torch.py
View File

@@ -101,8 +101,9 @@ class HFTorchInferenceModel(HFInferenceModel):
ret = super().set_input_parameters(parameters) ret = super().set_input_parameters(parameters)
# Hook onto input param setting for setting breakmodel stuff # Hook onto input param setting for setting breakmodel stuff
self.breakmodel_config.gpu_blocks = self.layers if self.breakmodel:
self.breakmodel_config.disk_blocks = self.disk_layers self.breakmodel_config.gpu_blocks = self.layers
self.breakmodel_config.disk_blocks = self.disk_layers
return ret return ret
@@ -303,20 +304,19 @@ class HFTorchInferenceModel(HFInferenceModel):
# Try to determine model type from either AutoModel or falling back to legacy # Try to determine model type from either AutoModel or falling back to legacy
try: try:
with lazy_loader.use_lazy_load(dematerialized_modules=True): if self.lazy_load:
metamodel = AutoModelForCausalLM.from_config(self.model_config) with lazy_loader.use_lazy_load(dematerialized_modules=True):
device_map = self.breakmodel_config.get_device_map(metamodel) metamodel = AutoModelForCausalLM.from_config(self.model_config)
tf_kwargs["device_map"] = self.breakmodel_config.get_device_map(metamodel)
print("Rodger rodger", tf_kwargs)
with lazy_loader.use_lazy_load( with lazy_loader.use_lazy_load(
enable=True, enable=self.lazy_load,
# DO NOT DEMATERIALIZE MODULES / INIT WEIGHTS EMPTY!!! IT WILL EXPLODE!!!!!!! # DO NOT DEMATERIALIZE MODULES / INIT WEIGHTS EMPTY!!! IT WILL EXPLODE!!!!!!!
dematerialized_modules=False, dematerialized_modules=False,
): ):
print(device_map)
model = AutoModelForCausalLM.from_pretrained( model = AutoModelForCausalLM.from_pretrained(
location, location,
# device_map="auto",
device_map=device_map,
offload_folder="accelerate-disk-cache", offload_folder="accelerate-disk-cache",
torch_dtype=torch.float16, torch_dtype=torch.float16,
**tf_kwargs, **tf_kwargs,