biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix][v1] Fix step pooler implementation and step pooling usage in v1 (#19956)

Browse Source 
Signed-off-by: Isotr0py <2037008807@qq.com>
This commit is contained in:
Isotr0py
2025-06-24 02:38:06 +08:00
committed by GitHub
parent 68aaeb3749
commit 61f4fc5dc6
14 changed files with 164 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
tests/models/multimodal/pooling/test_dse_qwen2_vl.py
View File

@@ -98,7 +98,7 @@ def _run_test(
max_model_len=8192) as vllm_model:
tokenizer = vllm_model.model.get_tokenizer()
texts = [
# this is necessary because vllm_model.encode will not apply any
# this is necessary because vllm_model.embed will not apply any
# templating to the prompt, and therefore lacks an image_pad
# token unless one is inserted beforehand (the (28,28) image
# above is converted to an image pad token by the chat template).
@@ -109,7 +109,7 @@ def _run_test(
# vllm will replace the pad token with the actual image,
# which may be a placeholder image, later.
]
vllm_outputs = vllm_model.encode(texts, images=input_images)
vllm_outputs = vllm_model.embed(texts, images=input_images)
hf_outputs = []
with hf_runner(model,