[Model Runner v2] Enable BGE M3 pooling embed token_classify (#50661)

Signed-off-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>
This commit is contained in:
Taneem Ibrahim
2026-08-02 12:46:03 -04:00
committed by GitHub
parent 96add737ff
commit 55c98e370a
2 changed files with 29 additions and 9 deletions
+6 -2
View File
@@ -18,9 +18,13 @@ from vllm.v1.worker.gpu.input_batch import InputBatch
from vllm.v1.worker.gpu.states import RequestState
_SUPPORTED_TASKS: frozenset[PoolingTask] = frozenset(
{"embed", "classify", "token_embed", "token_classify"}
{"embed", "classify", "token_embed", "token_classify", "embed&token_classify"}
)
# `embed&token_classify` concatenates a fixed-size embedding with per-token
# weights, so its output length scales with the prompt like the others here.
_TOKEN_TASKS: frozenset[PoolingTask] = frozenset(
{"token_embed", "token_classify", "embed&token_classify"}
)
_TOKEN_TASKS: frozenset[PoolingTask] = frozenset({"token_embed", "token_classify"})
class PoolingRunner: