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

optimize npu llama2 first token performance (#11451)

Browse source
This commit is contained in:
Yishuo Wang 2024-06-27 17:37:33 +08:00 committed by GitHub
parent 4e4ecd5095
commit 029ff15d28
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 19 additions and 7 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

26
python/llm/src/ipex_llm/transformers/npu_models/llama.py
View file

@ -219,13 +219,25 @@ def llama_attention_forward(
else: else:
causal_mask = None causal_mask = None
attn_output = torch.nn.functional.scaled_dot_product_attention( if query_states.size(2) == key_states.size(2):
query_states, # first token
key_states, from intel_npu_acceleration_library.functional import scaled_dot_product_attention
value_states, attn_output = scaled_dot_product_attention(
attn_mask=causal_mask, query_states,
is_causal=self.is_causal and attention_mask is None and q_len > 1, key_states,
) value_states,
attn_mask=causal_mask,
is_causal=self.is_causal and causal_mask is None and q_len > 1,
)
else:
# second+ token
attn_output = torch.nn.functional.scaled_dot_product_attention(
query_states,
key_states,
value_states,
attn_mask=causal_mask,
is_causal=self.is_causal and causal_mask is None and q_len > 1,
)
attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.transpose(1, 2).contiguous()