Support Roberta embedding models (#9387)

Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Signed-off-by: Flavia Beo <flavia.beo@ibm.com>
Co-authored-by: Flavia Beo <flavia.beo@ibm.com>
This commit is contained in:
Maximilien de Bayser
2024-11-14 21:23:29 +00:00
committed by GitHub
co-authored by Flavia Beo
parent 1dbae0329c
commit 4a18fd14ba
10 changed files with 202 additions and 14 deletions
+3
View File
@@ -98,6 +98,9 @@ void paged_attention_v1_launcher(
// NOTE(woosuk): To reduce the compilation time, we only compile for the
// head sizes that we use in the model. However, we can easily extend this
// to support any head size which is a multiple of 16.
case 32:
LAUNCH_PAGED_ATTENTION_V1(32);
break;
case 64:
LAUNCH_PAGED_ATTENTION_V1(64);
break;
+3
View File
@@ -104,6 +104,9 @@ void paged_attention_v2_launcher(
// NOTE(woosuk): To reduce the compilation time, we only compile for the
// head sizes that we use in the model. However, we can easily extend this
// to support any head size which is a multiple of 16.
case 32:
LAUNCH_PAGED_ATTENTION_V2(32);
break;
case 64:
LAUNCH_PAGED_ATTENTION_V2(64);
break;
+6
View File
@@ -385,6 +385,9 @@ void paged_attention_v1_impl_launcher(
int* seq_lens_ptr = seq_lens.data_ptr<int>();
switch (head_size) {
case 32:
LAUNCH_V1_ATTENTION_KERNEL(T, 32, BLOCK_SIZE);
break;
case 64:
LAUNCH_V1_ATTENTION_KERNEL(T, 64, BLOCK_SIZE);
break;
@@ -702,6 +705,9 @@ void paged_attention_v2_impl_launcher(
int* seq_lens_ptr = seq_lens.data_ptr<int>();
switch (head_size) {
case 32:
LAUNCH_V2_ATTENTION_KERNEL(T, 32, BLOCK_SIZE);
break;
case 64:
LAUNCH_V2_ATTENTION_KERNEL(T, 64, BLOCK_SIZE);
break;