rixty/KoboldAI-Client
1
1
Fork 0
mirror of https://github.com/KoboldAI/KoboldAI-Client.git synced 2024-12-12 08:36:28 +01:00
Code Issues Projects Releases Wiki Activity

Fix "is on the meta device" error when loading model with disk cache

Browse Source
This commit is contained in:
vfbd 2022-10-26 16:00:45 -04:00
parent 8ee795055c
commit 3233e78c56
1 changed files with 10 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
aiserver.py
View File

@ -2402,6 +2402,15 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal
if utils.num_shards is None or utils.current_shard >= utils.num_shards: if utils.num_shards is None or utils.current_shard >= utils.num_shards:
if utils.offload_index: if utils.offload_index:
for name, tensor in utils.named_buffers: for name, tensor in utils.named_buffers:
dtype = tensor.dtype
if convert_to_float16 and breakmodel.primary_device != "cpu" and vars.hascuda and (vars.breakmodel or vars.usegpu):
dtype = torch.float16
if breakmodel.primary_device == "cpu" or (not vars.usegpu and not vars.breakmodel):
dtype = torch.float32
if name in model_dict and model_dict[name].dtype is not dtype:
model_dict[name] = model_dict[name].to(dtype)
if tensor.dtype is not dtype:
tensor = tensor.to(dtype)
if name not in utils.offload_index: if name not in utils.offload_index:
accelerate.utils.offload_weight(tensor, name, "accelerate-disk-cache", index=utils.offload_index) accelerate.utils.offload_weight(tensor, name, "accelerate-disk-cache", index=utils.offload_index)
accelerate.utils.save_offload_index(utils.offload_index, "accelerate-disk-cache") accelerate.utils.save_offload_index(utils.offload_index, "accelerate-disk-cache")
@ -2574,7 +2583,7 @@ def load_model(use_gpu=True, gpu_layers=None, disk_layers=None, initial_load=Fal
if not args.colab or args.savemodel: if not args.colab or args.savemodel:
import shutil import shutil
tokenizer.save_pretrained("models/{}".format(vars.model.replace('/', '_'))) tokenizer.save_pretrained("models/{}".format(vars.model.replace('/', '_')))
if(vars.fp32_model): # Use save_pretrained to convert fp32 models to fp16 if(vars.fp32_model and ("breakmodel" not in globals() or not breakmodel.disk_blocks)): # Use save_pretrained to convert fp32 models to fp16, unless we are using disk cache because save_pretrained is not supported in that case
model = model.half() model = model.half()
model.save_pretrained("models/{}".format(vars.model.replace('/', '_')), max_shard_size="500MiB") model.save_pretrained("models/{}".format(vars.model.replace('/', '_')), max_shard_size="500MiB")
else: # For fp16 models, we can just copy the model files directly else: # For fp16 models, we can just copy the model files directly