test: stand up the config tiers two unit tests read from (#33294)
This commit is contained in:
@@ -552,10 +552,9 @@ def _acknowledge_deferred_cuda_ipc_cache_hits(
|
||||
parallel = get_parallel()
|
||||
if parallel.attn_tp_rank != 0:
|
||||
return
|
||||
server_args = get_server_args()
|
||||
# The pool's recycler uses ServerArgs.tp_size, so its acknowledgement must
|
||||
# The pool's recycler counts the whole TP group, so the acknowledgement must
|
||||
# match that count even when an attention subgroup is smaller.
|
||||
consumer_count = max(getattr(server_args, "tp_size", parallel.attn_tp_size), 1)
|
||||
consumer_count = max(get_server_args().tp_size, 1)
|
||||
for item in items:
|
||||
item.acknowledge_deferred_cuda_ipc_feature(consumer_count)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user