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

LLM: optimize CPU speculative decoding of chatglm3 (#9928)

Browse source 
* update

* fix style

* meet code review
This commit is contained in:
Ruonan Wang 2024-01-19 14:10:22 +08:00 committed by GitHub
parent 967714bac8
commit bf37b3a670
1 changed files with 6 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/models/chatglm2.py
View file

@ -367,17 +367,17 @@ def chatglm2_attention_forward_8eb45c(
def core_attn_forward_8eb45c(self, query_layer, key_layer, value_layer, attention_mask): def core_attn_forward_8eb45c(self, query_layer, key_layer, value_layer, attention_mask):
pytorch_major_version = int(torch.__version__.split('.')[0]) pytorch_major_version = int(torch.__version__.split('.')[0])
if pytorch_major_version >= 2 and (query_layer.device.type == 'xpu' or query_layer.size(0) > 1): if pytorch_major_version >= 2:
query_layer = query_layer.permute(1, 2, 0, 3) query_layer = query_layer.permute(1, 2, 0, 3)
L, S = query_layer.shape[2], key_layer.shape[2] L, S = query_layer.shape[2], key_layer.shape[2]
if attention_mask is None and L == S: if attention_mask is None and L == S:
context_layer = torch.nn.functional.scaled_dot_product_attention(query_layer, context_layer = F.scaled_dot_product_attention(query_layer.to(key_layer.dtype),
key_layer, key_layer,
value_layer, value_layer,
is_causal=True) is_causal=True)
else: else:
head_dim = query_layer.size(-1) head_dim = query_layer.size(-1)
attn = torch.matmul(query_layer, attn = torch.matmul(query_layer.to(key_layer.dtype),
key_layer.transpose(2, 3)) / math.sqrt(head_dim) key_layer.transpose(2, 3)) / math.sqrt(head_dim)
if attention_mask is not None: if attention_mask is not None:
attn_bias = torch.zeros(attention_mask.shape, dtype=query_layer.dtype, attn_bias = torch.zeros(attention_mask.shape, dtype=query_layer.dtype,