Skip to content

Flash attention enhancement of repeatKV #626

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
regisss merged 1 commit into from
Jan 23, 2024
Merged

Flash attention enhancement of repeatKV #626

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 3 additions & 3 deletions optimum/habana/transformers/models/llama/modeling_llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -228,9 +228,6 @@ def pre_attn_forward(
else:
past_key_value = None

key_states = gaudi_llama_repeat_kv(key_states, self.num_key_value_groups)
value_states = gaudi_llama_repeat_kv(value_states, self.num_key_value_groups)

if use_flash_attention and FusedSDPA:
import habana_frameworks.torch.hpu as ht

Expand All @@ -246,6 +243,9 @@ def pre_attn_forward(
attn_output = FusedSDPA.apply(query_states, key_states, value_states, None, 0.0, True, None)

else:
key_states = gaudi_llama_repeat_kv(key_states, self.num_key_value_groups)
value_states = gaudi_llama_repeat_kv(value_states, self.num_key_value_groups)

attn_weights = self.matmul_qk(query_states, key_states.transpose(2, 3)) * self.norm_factor

if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
Expand Down