[Pooling] Honor max_embed_len for chunked embeddings (#55551)

Signed-off-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>
This commit is contained in:
Taneem Ibrahim
2026-09-07 09:31:40 -04:00
committed by GitHub
parent 167858ea17
commit ecd600d91e
+1 -1
View File
@@ -179,7 +179,7 @@ class EmbeddingTokenizeParamsMixin(PoolingTokenizeParamsMixin):
pooler_config = model_config.pooler_config
if pooler_config is not None:
if pooler_config.enable_chunked_processing:
max_total_tokens = None
max_total_tokens = pooler_config.max_embed_len
else:
max_embed_len = pooler_config.max_embed_len or default_max_total_tokens
max_output_tokens = default_max_total_tokens - max_embed_len