ayo/ipex-llm
1
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 1

[LLM] Fix the model.device problem when cpu_embedding=True (#9971)

Browse source 
* Overwrite the device attribute for CPUPinnedParam

* Expose cpu_embedding=True for Linux users

* Fix python style
This commit is contained in:
Yuwen Hu 2024-01-23 18:51:11 +08:00 committed by GitHub
parent f82782cd3b
commit 8d28aa8e2b
2 changed files with 26 additions and 11 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

21
python/llm/src/bigdl/llm/transformers/convert.py
View file

@ -303,17 +303,16 @@ def _replace_with_low_bit_linear(model, qtype, modules_to_not_convert=None,
module.weight = None module.weight = None
elif cpu_embedding and type(module) == nn.Embedding: elif cpu_embedding and type(module) == nn.Embedding:
# skip user-defined Embedding layer # skip user-defined Embedding layer
if platform.system().lower() == 'windows': model._modules[name] = LLMEmbedding(
model._modules[name] = LLMEmbedding( num_embeddings=module.num_embeddings,
num_embeddings=module.num_embeddings, embedding_dim=module.embedding_dim,
embedding_dim=module.embedding_dim, padding_idx=module.padding_idx,
padding_idx=module.padding_idx, max_norm=module.max_norm,
max_norm=module.max_norm, norm_type=module.norm_type,
norm_type=module.norm_type, scale_grad_by_freq=module.scale_grad_by_freq,
scale_grad_by_freq=module.scale_grad_by_freq, sparse=module.sparse,
sparse=module.sparse, _weight=module.weight.data,
_weight=module.weight.data, )
)
# Remove the last key for recursion # Remove the last key for recursion
if len(list(module.children())) > 0: if len(list(module.children())) > 0:

16
python/llm/src/bigdl/llm/transformers/embedding.py
View file

@ -25,11 +25,27 @@ from typing import Optional
# To prevent insufficient available memory when moving embedding from XPU back to CPU, # To prevent insufficient available memory when moving embedding from XPU back to CPU,
# we can pin the embedding to CPU if `cpu_embedding==True`. # we can pin the embedding to CPU if `cpu_embedding==True`.
class CPUPinnedParam(Parameter): class CPUPinnedParam(Parameter):
# Overwrite the device attribute for CPUPinnedParam so that its device will be same as
# the device for model.to(device);
# With this device attribute, model.device will be same as the
# the device for model.to(device) even with cpu_embedding==True
@property
def device(self):
try:
return self._device
except AttributeError:
return super().device
@device.setter
def device(self, to_device):
self._device = to_device
def to(self, *args, **kwargs): def to(self, *args, **kwargs):
device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(*args, **kwargs) device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(*args, **kwargs)
if device is None: if device is None:
return super().to(*args, **kwargs) return super().to(*args, **kwargs)
elif device.type == 'xpu': elif device.type == 'xpu':
self.device = device
if convert_to_format is not None and self.dim() in (4, 5): if convert_to_format is not None and self.dim() in (4, 5):
return super().to('cpu', dtype, return super().to('cpu', dtype,
non_blocking, memory_format=convert_to_format) non_blocking, memory_format=convert_to_format)