mirror of
https://github.com/vllm-project/vllm.git
synced 2025-10-20 14:53:52 +08:00
[Bugfix][TPU] Fix megacore setting for v5e-litepod (#6397)
This commit is contained in:
@ -116,7 +116,7 @@ class PallasAttentionBackendImpl(AttentionImpl):
|
||||
|
||||
self.megacore_mode = None
|
||||
tpu_type = torch_xla.tpu.get_tpu_env()["TYPE"].lower()
|
||||
if not tpu_type.endswith("lite"):
|
||||
if "lite" not in tpu_type:
|
||||
if self.num_kv_heads % 2 == 0:
|
||||
self.megacore_mode = "kv_head"
|
||||
else:
|
||||
|
Reference in New Issue
Block a user