A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 6, 05:45 UTC
always use `embed&token_classify` for bge-m3 (#37632)
Signed-off-by: augusto.yjh <augusto.yjh@antgroup.com> Co-authored-by: wang.yuqi <yuqi.wang@daocloud.io>
Augusto Yao committed
Mar 23, 2026 at 11:10 UTC
6e04e79326109dfcbd86268714b779c0af3698b7
1 file changed
+6
-19
tests/plugins/bge_m3_sparse_plugin/bge_m3_sparse_processor/sparse_embeddings_processor.py
+6
-19
@@ -65,21 +65,10 @@ class BgeM3SparseEmbeddingsProcessor(
65
f"Unsupported task {raw_embed_request}, "
66
f"Supported tasks are {EMBED_TASKS}"
67
)
68
- has_dense_embed = True
69
- if raw_embed_request.embed_task == "dense":
70
- params.task = "embed"
71
- params.skip_reading_prefix_cache = False
72
- elif raw_embed_request.embed_task == "sparse":
73
- params.task = "token_classify"
74
- has_dense_embed = False
75
- else:
76
- params.task = "embed&token_classify"
68
+ params.task = "embed&token_classify"
69
params.use_activation = raw_embed_request.use_activation
70
if params.use_activation is None:
71
params.use_activation = True
80
- if not has_dense_embed:
81
- params.dimensions = None
82
- return params
72
73
params.dimensions = raw_embed_request.dimensions
74
@@ -170,13 +159,11 @@ class BgeM3SparseEmbeddingsProcessor(
159
raw_request = self._get_sparse_embedding_request(request_id)
160
has_dense_embed = raw_request.embed_task in ["dense", "dense&sparse"]
161
has_sparse_embed = raw_request.embed_task in ["sparse", "dense&sparse"]
173
- embed_dimensions = 0
174
- if has_dense_embed:
175
- embed_dimensions = (
176
- self.embed_dimensions
177
- if raw_request.dimensions is None
178
- else raw_request.dimensions
179
- )
162
+ embed_dimensions = (
163
+ self.embed_dimensions
164
+ if raw_request.dimensions is None
165
+ else raw_request.dimensions
166
+ )
167
for idx in range(len(model_output)):
168
mo = model_output[idx]
169
sparse_embedding_dict: dict[int, float] = {}