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

LLM: add fuse rope and norm optimization for Baichuan. (#9166)

Browse source 
* add fuse rope optimization.

* add rms norm optimization.
This commit is contained in:
Cengguang Zhang 2023-10-13 17:36:52 +08:00 committed by GitHub
parent db7f938fdc
commit 51a133de56
3 changed files with 26 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
python/llm/src/bigdl/llm/transformers/convert.py
View file

@ -275,6 +275,9 @@ def optimize(model):
module.BaichuanAttention, module.BaichuanAttention,
baichuan_attention_forward_13b baichuan_attention_forward_13b
) )
convert_forward(model,
module.RMSNorm,
llama_rms_norm_forward)
elif model.config.model_type == "baichuan": elif model.config.model_type == "baichuan":
# baichuan1 # baichuan1
@ -296,6 +299,9 @@ def optimize(model):
module.BaichuanAttention, module.BaichuanAttention,
baichuan_attention_forward_13b baichuan_attention_forward_13b
) )
convert_forward(model,
module.RMSNorm,
llama_rms_norm_forward)
elif model.config.model_type == "gpt_neox": elif model.config.model_type == "gpt_neox":
from bigdl.llm.transformers.models.gptneox import gptneox_attention_forward from bigdl.llm.transformers.models.gptneox import gptneox_attention_forward

13
python/llm/src/bigdl/llm/transformers/models/baichuan.py
View file

@ -28,6 +28,7 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from bigdl.llm.utils.common import invalidInputError from bigdl.llm.utils.common import invalidInputError
from bigdl.llm.transformers.models.utils import init_kv_cache, extend_kv_cache, append_kv_cache from bigdl.llm.transformers.models.utils import init_kv_cache, extend_kv_cache, append_kv_cache
from bigdl.llm.transformers.models.utils import rotate_half, apply_rotary_pos_emb from bigdl.llm.transformers.models.utils import rotate_half, apply_rotary_pos_emb
from bigdl.llm.transformers.models.utils import apply_rotary_pos_emb_no_cache_xpu
KV_CACHE_ALLOC_BLOCK_LENGTH = 256 KV_CACHE_ALLOC_BLOCK_LENGTH = 256
@ -56,9 +57,15 @@ def baichuan_attention_forward_7b(
kv_seq_len = key_states.shape[-2] kv_seq_len = key_states.shape[-2]
if past_key_value is not None: if past_key_value is not None:
kv_seq_len += past_key_value[0].shape[-2] kv_seq_len += past_key_value[0].shape[-2]
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) if query_states.device.type == "xpu" and not (self.training and query_states.requires_grad):
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, query_states, key_states = apply_rotary_pos_emb_no_cache_xpu(query_states,
cos, sin, position_ids, "baichuan") key_states,
position_ids,
"baichuan")
else:
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
query_states, key_states = apply_rotary_pos_emb(query_states, key_states,
cos, sin, position_ids, "baichuan")
# [bsz, nh, t, hd] # [bsz, nh, t, hd]
# if past_key_value is not None: # if past_key_value is not None:

13
python/llm/src/bigdl/llm/transformers/models/baichuan2.py
View file

@ -28,6 +28,7 @@ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from bigdl.llm.utils.common import invalidInputError from bigdl.llm.utils.common import invalidInputError
from bigdl.llm.transformers.models.utils import init_kv_cache, extend_kv_cache, append_kv_cache from bigdl.llm.transformers.models.utils import init_kv_cache, extend_kv_cache, append_kv_cache
from bigdl.llm.transformers.models.utils import rotate_half, apply_rotary_pos_emb from bigdl.llm.transformers.models.utils import rotate_half, apply_rotary_pos_emb
from bigdl.llm.transformers.models.utils import apply_rotary_pos_emb_no_cache_xpu
from transformers.utils import logging, ContextManagers from transformers.utils import logging, ContextManagers
logger = logging.get_logger(__name__) logger = logging.get_logger(__name__)
@ -68,9 +69,15 @@ def baichuan_attention_forward_7b(
kv_seq_len = key_states.shape[-2] kv_seq_len = key_states.shape[-2]
if past_key_value is not None: if past_key_value is not None:
kv_seq_len += past_key_value[0].shape[-2] kv_seq_len += past_key_value[0].shape[-2]
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) if query_states.device.type == "xpu" and not (self.training and query_states.requires_grad):
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, query_states, key_states = apply_rotary_pos_emb_no_cache_xpu(query_states,
cos, sin, position_ids, "baichuan") key_states,
position_ids,
"baichuan")
else:
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
query_states, key_states = apply_rotary_pos_emb(query_states, key_states,
cos, sin, position_ids, "baichuan")
# [bsz, nh, t, hd] # [bsz, nh, t, hd]
# if past_key_value is not None: # if past_key_value is not None: