[Bugfix][Frontend] Preserve caller parameters in offline pooling (#58842)
Signed-off-by: 子华 <huaxi.shx@alibaba-inc.com> Co-authored-by: Codex <noreply@openai.com>
shaohuaxi committed
Sep 28, 2026 at 22:02 UTC
36768d1bfd39094681cdbc8cb37d4b31c0729c89
2 files changed
+5
-2
vllm/entrypoints/pooling/base/io_processor.py
+4
-1
@@ -207,7 +207,10 @@ class PoolingIOProcessor:
207
else:
208
pooling_params = ctx.pooling_params
209
210
- params_seq = self._params_to_seq(pooling_params, num_requests)
210
+ # Clone to avoid modifying the caller's pooling parameters.
211
+ params_seq = [
212
+ param.clone() for param in self._params_to_seq(pooling_params, num_requests)
213
+ ]
214
215
for param in params_seq:
216
if param.task is None:
vllm/entrypoints/pooling/pooling/io_processor.py
+1
-1
@@ -168,7 +168,7 @@ class PluginWithIOProcessorPlugins(PoolingIOProcessor):
168
pooling_params = ctx.pooling_params
169
170
params_seq: list[PoolingParams] = [
171
- self.io_processor.merge_pooling_params(param)
171
+ self.io_processor.merge_pooling_params(param.clone())
172
for param in self._params_to_seq(
173
pooling_params,
174
num_requests,