[Bugfix] Skip PP sampled-token receive on last rank during async scheduling (#40749)

Signed-off-by: Adam Winstanley <adam@winstanley.industries>
This commit is contained in:
wi-adam
2026-05-06 05:31:14 +00:00
committed by GitHub
parent 2d7d6cf765
commit b53c507bc9
2 changed files with 54 additions and 1 deletions
+1 -1
View File
@@ -4182,7 +4182,7 @@ class GPUModelRunner(
kv_connector_output = self.kv_connector_output
self.kv_connector_output = None
# receive sampled token ids from the last PP rank.
if self.use_async_scheduling and get_pp_group().world_size > 1:
if self.use_async_scheduling and not get_pp_group().is_last_rank:
self._pp_receive_prev_sampled_token_ids_to_input_batch()
if not kv_connector_output:
return None # type: ignore[return-value]