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

Hotfix of BCE-Emdedding model (#12490)

Browse source
This commit is contained in:
binbin Deng 2024-12-03 18:16:04 +08:00 committed by GitHub
parent 80f15e41f5
commit c59284418c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 3 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
python/llm/src/ipex_llm/transformers/npu_model.py
View file

@ -766,7 +766,7 @@ class EmbeddingModel(_BaseAutoModelClass):
optimize_llm_pre(model, qtype, mixed_precision, optimize_llm_pre(model, qtype, mixed_precision,
quantization_group_size=quantization_group_size) quantization_group_size=quantization_group_size)
cls.load_convert_fp16(qtype, model.encoder, "cpu", modules_to_not_convert, cls.load_convert_fp16(qtype, model.encoder, "cpu", modules_to_not_convert,
quantization_group_size, None, *args, **kwargs) quantization_group_size)
create_npu_kernels(model.encoder) create_npu_kernels(model.encoder)
model = model.eval() model = model.eval()
logger.info(f"Finish to convert model") logger.info(f"Finish to convert model")
@ -781,7 +781,7 @@ class EmbeddingModel(_BaseAutoModelClass):
@classmethod @classmethod
def load_convert_fp16(cls, q_k, optimize_model, device, modules_to_not_convert, def load_convert_fp16(cls, q_k, optimize_model, device, modules_to_not_convert,
group_size=0, imatrix_data=None, *arg, **kwarg): group_size=0, imatrix_data=None):
from ipex_llm.transformers.npu_models.xlm_mp import replace_with_FP16Linear from ipex_llm.transformers.npu_models.xlm_mp import replace_with_FP16Linear
replace_with_FP16Linear(optimize_model, q_k, device=device, replace_with_FP16Linear(optimize_model, q_k, device=device,
modules_to_not_convert=modules_to_not_convert, modules_to_not_convert=modules_to_not_convert,

2
python/llm/src/ipex_llm/transformers/npu_models/xlm_mp.py
View file

@ -711,7 +711,7 @@ class XLMLayerNorm(torch.nn.Module):
@module_optimization @module_optimization
def replace_with_Layernorm(layer, qtype=None, device='NPU', def replace_with_Layernorm(layer, qtype=None, device='NPU',
modules_to_not_convert=[], group_size=0): modules_to_not_convert=[], group_size=0, **kwargs):
if isinstance(layer, torch.nn.LayerNorm): if isinstance(layer, torch.nn.LayerNorm):
return XLMLayerNorm( return XLMLayerNorm(
weight=layer.weight.to(torch.float16), weight=layer.weight.to(torch.float16),