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

[NPU] further fix of qwen2 int8 pipeline & C++ (#12449)

Browse source 
* fix

* fix style
This commit is contained in:
Ruonan Wang 2024-11-26 00:39:39 -08:00 committed by GitHub
parent 303b104c10
commit 24b46b2b19
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 8 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
python/llm/src/ipex_llm/transformers/npu_model.py
View file

@ -231,7 +231,7 @@ class _BaseAutoModelClass:
from intel_npu_acceleration_library.compiler import create_npu_kernels from intel_npu_acceleration_library.compiler import create_npu_kernels
model = kwargs.pop("model") model = kwargs.pop("model")
qtype = kwargs.pop("qtype", "sym_int4") qtype = kwargs.pop("qtype", "sym_int4_rtn")
mixed_precision = kwargs.pop("mixed_precision", False) mixed_precision = kwargs.pop("mixed_precision", False)
quantization_group_size = kwargs.pop("quantization_group_size", 0) quantization_group_size = kwargs.pop("quantization_group_size", 0)
modules_to_not_convert = kwargs.pop("modules_to_not_convert", []) modules_to_not_convert = kwargs.pop("modules_to_not_convert", [])
@ -280,6 +280,7 @@ class _BaseAutoModelClass:
max_prompt_len=max_prompt_len, max_prompt_len=max_prompt_len,
transpose_value_cache=transpose_value_cache, transpose_value_cache=transpose_value_cache,
group_size=quantization_group_size, group_size=quantization_group_size,
qtype=qtype,
convert_model=convert_model, convert_model=convert_model,
save_directory=save_directory) save_directory=save_directory)
model.save_low_bit = types.MethodType(save_low_bit, model) model.save_low_bit = types.MethodType(save_low_bit, model)

7
python/llm/src/ipex_llm/transformers/npu_pipeline_model/convert_pipeline.py
View file

@ -193,13 +193,18 @@ def convert_llm(model: torch.nn.Module,
max_prompt_len: int, max_prompt_len: int,
transpose_value_cache: bool, transpose_value_cache: bool,
group_size: int, group_size: int,
qtype: str,
convert_model: bool=False, convert_model: bool=False,
save_directory: str=None): save_directory: str=None):
# whether to set layernorm weight as const # whether to set layernorm weight as const
layernorm_const = os.environ.get("IPEX_LLM_LAYERNORM_CONST", "1") == "1" layernorm_const = os.environ.get("IPEX_LLM_LAYERNORM_CONST", "1") == "1"
if group_size == 0: if group_size == 0:
n_splits_linear = 1 n_splits_linear = 1
n_splits_down_proj = 2 if model.config.intermediate_size == 18944 else 1 if qtype == "sym_int8_rtn":
# do not split mlp down_proj for Qwen2-7B & sym_int8
n_splits_down_proj = 1
else:
n_splits_down_proj = 2 if model.config.intermediate_size == 18944 else 1
else: else:
n_splits_linear = model.config.hidden_size // group_size n_splits_linear = model.config.hidden_size // group_size
n_splits_down_proj = model.config.intermediate_size // group_size n_splits_down_proj = model.config.intermediate_size // group_size