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

optimize internlm2 xcomposer agin (#11124)

Browse source
This commit is contained in:
Yishuo Wang 2024-05-24 13:44:52 +08:00 committed by GitHub
parent 9372ce87ce
commit 1db9d9a63b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 14 additions and 6 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

18
python/llm/src/ipex_llm/transformers/models/internlm.py
View file

@ -45,8 +45,9 @@ from torch import nn
from ipex_llm.utils.common import invalidInputError from ipex_llm.utils.common import invalidInputError
from ipex_llm.transformers.models.utils import init_kv_cache, extend_kv_cache, \ from ipex_llm.transformers.models.utils import init_kv_cache, extend_kv_cache, \
append_kv_cache, is_enough_kv_cache_room_4_31 append_kv_cache, is_enough_kv_cache_room_4_31
from ipex_llm.transformers.models.utils import apply_rotary_pos_emb from ipex_llm.transformers.models.utils import should_use_fuse_rope, apply_rotary_pos_emb
from ipex_llm.transformers.models.utils import apply_rotary_pos_emb_no_cache_xpu from ipex_llm.transformers.models.utils import apply_rotary_pos_emb_no_cache_xpu
from ipex_llm.transformers.models.utils import apply_rotary_pos_emb_cache_freq_xpu
from ipex_llm.transformers.models.utils import use_quantize_kv_cache, restore_fp8_kv_cache from ipex_llm.transformers.models.utils import use_quantize_kv_cache, restore_fp8_kv_cache
from ipex_llm.transformers.models.utils import update_past_key_value from ipex_llm.transformers.models.utils import update_past_key_value
from ipex_llm.transformers.models.utils import use_sdp, use_sdp_causal from ipex_llm.transformers.models.utils import use_sdp, use_sdp_causal
@ -83,7 +84,7 @@ def internlm_attention_forward(
if past_key_value is not None: if past_key_value is not None:
enough_kv_room = is_enough_kv_cache_room_4_31(past_key_value, seq_len=kv_seq_len) enough_kv_room = is_enough_kv_cache_room_4_31(past_key_value, seq_len=kv_seq_len)
kv_seq_len += past_key_value[0].shape[-2] kv_seq_len += past_key_value[0].shape[-2]
if query_states.device.type == "xpu" and not (self.training and query_states.requires_grad): if should_use_fuse_rope(hidden_states, position_ids, self.training):
query_states, key_states = apply_rotary_pos_emb_no_cache_xpu(query_states, query_states, key_states = apply_rotary_pos_emb_no_cache_xpu(query_states,
key_states, key_states,
position_ids, position_ids,
@ -228,7 +229,7 @@ def internlm2_attention_forward(
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]
if query_states.device.type == "xpu" and not (self.training and query_states.requires_grad): if should_use_fuse_rope(hidden_states, position_ids, self.training):
query_states, key_states = apply_rotary_pos_emb_no_cache_xpu(query_states, query_states, key_states = apply_rotary_pos_emb_no_cache_xpu(query_states,
key_states, key_states,
position_ids, position_ids,
@ -376,9 +377,16 @@ def internlm_xcomposser2_attention_forward(
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]
# IPEX-LLM OPT: fuse rope
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
query_states, key_states = apply_rotary_pos_emb( if should_use_fuse_rope(hidden_states, position_ids, self.training):
query_states, key_states, cos, sin, position_ids, "internlm") query_states, key_states = apply_rotary_pos_emb_cache_freq_xpu(
query_states, key_states, sin, cos, "internlm", position_ids
)
else:
query_states, key_states = apply_rotary_pos_emb(
query_states, key_states, cos, sin, position_ids, "internlm")
# IPEX-LLM OPT: kv cache and quantzie kv cache # IPEX-LLM OPT: kv cache and quantzie kv cache
use_quantize_kv = use_quantize_kv_cache(self.wqkv, hidden_states) use_quantize_kv = use_quantize_kv_cache(self.wqkv, hidden_states)

2
python/llm/src/ipex_llm/transformers/models/utils.py
View file

@ -233,7 +233,7 @@ def apply_rotary_pos_emb_cache_freq_xpu(q, k, sin, cos, model_family, position_i
k_embed = torch.empty(k.shape, dtype=k.dtype, device=k.device) k_embed = torch.empty(k.shape, dtype=k.dtype, device=k.device)
if model_family in ["qwen", "mixtral"]: if model_family in ["qwen", "mixtral"]:
linear_q4_0.apply_rotary_embedding_half_q_and_k_cache_freq(q, k, sin, cos, q_embed, k_embed) linear_q4_0.apply_rotary_embedding_half_q_and_k_cache_freq(q, k, sin, cos, q_embed, k_embed)
elif model_family in ["qwen2", "yuan", "stablelm", "qwen2_moe"]: elif model_family in ["qwen2", "yuan", "stablelm", "qwen2_moe", "internlm"]:
cos = cos.to(q.dtype) cos = cos.to(q.dtype)
sin = sin.to(q.dtype) sin = sin.to(q.dtype)
cos = cos.squeeze(1).squeeze(0) # [seq_len, dim] cos = cos.squeeze(1).squeeze(0) # [seq_len, dim]