[Bugfix][Multimodal] Encode text in mixed CLIP/SigLIP pooling batches (#53165)

Signed-off-by: Prudhvivuda <prudhvi12042001@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Prudhvi Vuda
2026-08-24 10:45:42 -07:00
committed by GitHub
co-authored by mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 4f686e182a
commit 342b8ebd8b
5 changed files with 154 additions and 11 deletions
@@ -52,6 +52,31 @@ def _run_test(
vllm_model.embed(texts)
vllm_model.embed([""], images=images)
# Mixed image+text batch must not skip the text encoder (#53091).
mixed_outputs = vllm_outputs_per_case[2]
check_embeddings_close(
embeddings_0_lst=[vllm_outputs_per_case[0][0]],
embeddings_1_lst=[mixed_outputs[0]],
name_0="text_only",
name_1="mixed_text",
)
check_embeddings_close(
embeddings_0_lst=[vllm_outputs_per_case[1][0]],
embeddings_1_lst=[mixed_outputs[1]],
name_0="image_only",
name_1="mixed_image",
)
empty_text = vllm_model.embed([""])
empty_sim = torch.nn.functional.cosine_similarity(
torch.tensor(mixed_outputs[0]),
torch.tensor(empty_text[0]),
dim=0,
)
assert empty_sim < 0.99, (
"Mixed-batch text embedding collapsed to the empty-string vector "
f"(cosine={empty_sim:.4f})"
)
with hf_runner(model, dtype=dtype, auto_cls=CLIPModel) as hf_model:
hf_outputs_per_case = []
for input_texts, input_images in input_cases:
@@ -101,6 +126,7 @@ def test_models(
input_cases = [
(HF_TEXT_PROMPTS, text_images),
(HF_IMAGE_PROMPTS, images),
([HF_TEXT_PROMPTS[0], HF_IMAGE_PROMPTS[0]], [None, images[0]]),
]
_run_test(
@@ -0,0 +1,42 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
import torch
from vllm.model_executor.models.clip import (
dual_encoder_has_text_tokens,
merge_dual_encoder_text_and_vision,
)
pytestmark = [pytest.mark.cpu_test, pytest.mark.skip_global_cleanup]
def test_text_only_batch_runs_text_encoder():
assert dual_encoder_has_text_tokens(False, None)
assert dual_encoder_has_text_tokens(False, torch.zeros(4, dtype=torch.bool))
def test_vision_only_batch_skips_text_encoder():
is_multimodal = torch.ones(8, dtype=torch.bool)
assert not dual_encoder_has_text_tokens(True, is_multimodal)
def test_mixed_batch_runs_text_encoder():
# Image tokens followed by a text sequence: the old batch-wide flag
# treated this as vision-only and skipped the text encoder (#53091).
is_multimodal = torch.tensor([True, True, True, False, False, False])
assert dual_encoder_has_text_tokens(True, is_multimodal)
def test_missing_token_mask_with_mm_embeddings_is_vision_only():
assert not dual_encoder_has_text_tokens(True, None)
def test_merge_keeps_vision_on_mm_tokens():
text = torch.zeros(4, 2)
vision = torch.ones(4, 2)
is_multimodal = torch.tensor([True, True, False, False])
out = merge_dual_encoder_text_and_vision(text, vision, is_multimodal)
assert torch.equal(out[:2], vision[:2])
assert torch.equal(out[2:], text[2:])
@@ -63,6 +63,20 @@ def _run_test(
vllm_model.embed(texts)
vllm_model.embed([""], images=images)
mixed_outputs = vllm_outputs_per_case[2]
check_embeddings_close(
embeddings_0_lst=[vllm_outputs_per_case[0][0]],
embeddings_1_lst=[mixed_outputs[0]],
name_0="text_only",
name_1="mixed_text",
)
check_embeddings_close(
embeddings_0_lst=[vllm_outputs_per_case[1][0]],
embeddings_1_lst=[mixed_outputs[1]],
name_0="image_only",
name_1="mixed_image",
)
with hf_runner(model, dtype=dtype, auto_cls=SiglipModel) as hf_model:
hf_outputs_per_case = []
for input_texts, input_images, tokenization_kwargs in input_cases:
@@ -122,6 +136,14 @@ def test_models(
},
),
(HF_IMAGE_PROMPTS, images, {}),
(
[HF_TEXT_PROMPTS[0], HF_IMAGE_PROMPTS[0]],
[None, images[0]],
{
"padding": "max_length",
"max_length": 64,
},
),
]
_run_test(