mirror of
https://github.com/wassname/vllm.git
synced 2026-10-03 13:02:20 +08:00
[TPU] Update pallas.py to support trillium (#8871)
This commit is contained in:
1 parent
6d792d2f31
commit
8df2dc3c88
1 file changed
+1
-1
@@ -130,7 +130,7 @@ class PallasAttentionBackendImpl(AttentionImpl):
|
||||
assert tpu_type is not None
|
||||
tpu_type = tpu_type.lower()
|
||||
|
||||
if "lite" not in tpu_type:
|
||||
if (("lite" not in tpu_type) and ("v6" not in tpu_type)):
|
||||
if self.num_kv_heads % 2 == 0:
|
||||
self.megacore_mode = "kv_head"
|
||||
else:
|
||||
|
||||
Reference in new issue
Block a user