[Pooling] Support prompt embeddings in MRV2 decoder pooling (#56639)

Signed-off-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>
This commit is contained in:
Taneem Ibrahim
2026-09-13 05:49:45 +00:00
committed by GitHub
parent 93e051e290
commit 0cf266a469
2 changed files with 2 additions and 2 deletions
+2 -1
View File
@@ -64,7 +64,7 @@ class PoolingRunner:
req_id: str,
req_index: int,
pooling_params: PoolingParams,
prompt_token_ids: list[int],
prompt_token_ids: list[int] | None,
) -> None:
task = pooling_params.task
if task not in self.supported_tasks:
@@ -77,6 +77,7 @@ class PoolingRunner:
self.pooling_params[req_index] = pooling_params
self.pooling_states[req_index] = PoolingStates()
if pooling_params.requires_token_ids:
assert prompt_token_ids is not None
self.prompt_token_ids[req_index] = torch.tensor(
prompt_token_ids, dtype=torch.int64
)