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

optimize phi3 1st token performance (#10981)

Browse source
This commit is contained in:
Yishuo Wang 2024-05-10 17:33:46 +08:00 committed by GitHub
parent cfed76b2ed
commit ad96f32ce0
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 26 additions and 13 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

5
python/llm/src/ipex_llm/transformers/models/phi3.py
View file

@ -42,6 +42,7 @@ from ipex_llm.transformers.models.utils import (
from ipex_llm.transformers.models.utils import mlp_fusion_check, SILU
from ipex_llm.transformers.models.utils import use_new_esimd_sdp_fp16, use_quantize_kv_cache
from ipex_llm.transformers.models.utils import use_sdp_fp8, restore_fp8_kv_cache
from ipex_llm.transformers.models.utils import use_sdp_causal
from ipex_llm.transformers.kv import DynamicNormalCache, DynamicFp8Cache
from typing import Optional, Tuple, List
@ -148,6 +149,10 @@ def attention_forward(
if isinstance(past_key_value, DynamicFp8Cache):
key_states, value_states = restore_fp8_kv_cache(key_states, value_states,
query_states.dtype)
if use_sdp_causal(q_len, kv_seq_len, query_states, self.training):
import linear_q4_0
attn_output = linear_q4_0.sdp_causal(query_states, key_states, value_states)
else:
# repeat k/v heads if n_kv_heads < n_heads
key_states = repeat_kv(key_states, self.num_key_value_groups)
value_states = repeat_kv(value_states, self.num_key_value_groups)

8
python/llm/src/ipex_llm/transformers/models/utils.py
View file

@ -384,6 +384,14 @@ def use_sdp_fp8(q_len, k_len, query_states):
return True
def use_sdp_causal(q_len, kv_len, query_states, training):
return (
q_len == kv_len # first token
and query_states.device.type == "xpu" # GPU
and not query_states.requires_grad and not training # no training
)
def mlp_fusion_check(x, qtype, training):
invalidInputError(x.dim() == 2,
"Here input x's dim should be 2.")