huggingface/text-generation-inference
5
0
Fork 0
mirror of https://github.com/huggingface/text-generation-inference.git synced 2025-09-11 04:14:52 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat: avoid copy for partial rotary embeddings

Browse Source
This commit is contained in:
drbh 2024-01-23 15:42:10 +00:00
parent c49332adb6
commit 18f13a1b5f
1 changed files with 5 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
server/text_generation_server/models/custom_modeling/flash_phi_modeling.py
View File

@ -180,16 +180,11 @@ class FlashPhiAttention(torch.nn.Module):
query = query.view(-1, self.num_heads, self.head_size) query = query.view(-1, self.num_heads, self.head_size)
kv = kv.view(-1, 2, self.num_key_value_heads, self.head_size) kv = kv.view(-1, 2, self.num_key_value_heads, self.head_size)
# Apply partial rotary embedding and store the end of the embedding # Apply partial positional embeddings in place
query_pass = query[:, :, self.rotary_emb_dim:] self.rotary_emb(
key_pass = torch.select(kv, dim=1, index=0)[:, :, self.rotary_emb_dim:] query[:, :, :self.rotary_emb_dim], kv[:, 0, :, :self.rotary_emb_dim],
cos, sin
# Apply in place positional rotary embeddings )
self.rotary_emb(query, torch.select(kv, dim=1, index=0), cos, sin)
# Restore the query and key from the partial rotary embedding
kv[:, 0, :, self.rotary_emb_dim:] = key_pass
query[:, :, self.rotary_emb_dim:] = query_pass
# Reshape key and value and cache # Reshape key and value and cache
paged_attention.reshape_and_cache( paged_attention.reshape_and_cache(