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

Clean npu dtype branch (#11515)

Browse source 
* clean branch

* create_npu_kernels
This commit is contained in:
Zhao Changmin 2024-07-05 15:45:26 +08:00 committed by GitHub
parent 14ce058004
commit f7e957aaf9
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 27 additions and 51 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

5
python/llm/src/ipex_llm/transformers/model.py
View file

@ -332,6 +332,11 @@ class _BaseAutoModelClass:
else:
kwargs["pretraining_tp"] = 1
q_k = load_in_low_bit if load_in_low_bit else "sym_int4"
invalidInputError(q_k not in ["sym_int4_rtn", "sym_int8_rtn"],
f"The dtype {q_k} is specified for NPU"
"and cannot be used on CPU and GPU")
imatrix_file = kwargs.pop("imatrix", None)
if q_k in ["gguf_iq2_xxs", "gguf_iq2_xs", "gguf_iq1_s"]:
invalidInputError(imatrix_file is not None,

73
python/llm/src/ipex_llm/transformers/npu_model.py
View file

@ -25,8 +25,6 @@ from unittest.mock import patch
from transformers.dynamic_module_utils import get_imports
from transformers.configuration_utils import PretrainedConfig
import intel_npu_acceleration_library as npu_lib
from ipex_llm.utils.common.log4Error import invalidInputError
from ipex_llm.transformers.utils import logger
from ipex_llm.transformers.npu_models.convert import optimize_llm
@ -90,23 +88,12 @@ class _BaseAutoModelClass:
warnings.warn("`torch_dtype` will be ignored, `torch.float` will be used")
kwargs['torch_dtype'] = torch.float
low_bit = kwargs.pop('load_in_low_bit', 'fp32')
try:
# for intel_npu_acceleration_library >= 1.1.0
from intel_npu_acceleration_library.dtypes import int8, int4
qtype_map = {
'sym_int4': "sym_int4_rtn",
'sym_int8': "sym_int8_rtn",
'fp16': torch.half,
'fp32': torch.float,
}
except ImportError as _e:
# for intel_npu_acceleration_library < 1.1.0
qtype_map = {
'sym_int8': torch.int8,
'fp16': torch.half,
'fp32': torch.float,
}
low_bit = kwargs.pop('load_in_low_bit', 'sym_int4')
qtype_map = {
'sym_int4': "sym_int4_rtn",
'sym_int8': "sym_int8_rtn",
}
invalidInputError(low_bit in qtype_map.keys(),
f"unsupported low_bit: {low_bit}, "
f"only {list(qtype_map.keys())} are supported")
@ -143,22 +130,15 @@ class _BaseAutoModelClass:
model.config.update({"bigdl_lcmu_enabled": False})
logger.info(f"Converting model, it may takes up to several minutes ...")
try:
# for intel_npu_acceleration_library >= 1.1.0
from intel_npu_acceleration_library.quantization import quantize_model
from intel_npu_acceleration_library.compiler import create_npu_kernels
with torch.no_grad():
optimize_llm(model)
if qtype in ["sym_int8_rtn", "sym_int4_rtn"]:
cls.load_convert(qtype, model, 'cpu', *args, **kwargs)
else:
if not qtype.is_floating_point:
model = quantize_model(model, qtype)
create_npu_kernels(model)
model = model.eval()
except ImportError as _e:
# for intel_npu_acceleration_library < 1.1.0
model = npu_lib.compile(model, qtype, False)
from intel_npu_acceleration_library.compiler import create_npu_kernels
with torch.no_grad():
optimize_llm(model)
cls.load_convert(qtype, model, 'cpu', *args, **kwargs)
create_npu_kernels(model)
model = model.eval()
logger.info(f"Finish to convert model")
model.config.update({"bigdl_transformers_low_bit": qtype})
@ -313,22 +293,13 @@ class _BaseAutoModelClass:
# Loading args may differ based on their usage
quant_device = "meta" if bigdl_lcmu_enabled else "cpu"
logger.info(f"Converting model, it may takes up to several minutes ...")
try:
# for intel_npu_acceleration_library >= 1.1.0
from intel_npu_acceleration_library.quantization import quantize_model
from intel_npu_acceleration_library.compiler import create_npu_kernels
with torch.no_grad():
optimize_llm(model)
if qtype in ["sym_int8_rtn", "sym_int4_rtn"]:
cls.load_convert(qtype, model, quant_device, *model_args, **kwargs)
else:
if not qtype.is_floating_point:
model = quantize_model(model, qtype)
create_npu_kernels(model)
model = model.eval()
except ImportError as _e:
# for intel_npu_acceleration_library < 1.1.0
model = npu_lib.compile(model, qtype, False)
from intel_npu_acceleration_library.compiler import create_npu_kernels
with torch.no_grad():
optimize_llm(model)
cls.load_convert(qtype, model, quant_device, *model_args, **kwargs)
create_npu_kernels(model)
model = model.eval()
if is_sharded:
loaded_state_dict_keys = sharded_metadata["all_checkpoint_keys"]