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, 13:15 UTC
[CI] Fix PaddleOCR-VL HF test failure due to create_causal_mask API rename (#37328)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Andreas Karatzas committed
Mar 18, 2026 at 04:44 UTC
eaf7c9b976799c0d8e6b1ffd9bd4c0b6e74e988d
2 files changed
+26
tests/models/multimodal/generation/test_common.py
+1
@@ -777,6 +777,7 @@ VLM_TEST_SETTINGS = {
777
max_model_len=8192,
778
max_num_seqs=2,
779
auto_cls=AutoModelForCausalLM,
780
+ patch_hf_runner=model_utils.paddleocr_vl_patch_hf_runner,
781
image_size_factors=[(0.25,)],
782
marks=[
783
pytest.mark.skipif(
tests/models/multimodal/generation/vlm_utils/model_utils.py
+25
@@ -1149,6 +1149,31 @@ def ovis2_5_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
1149
return hf_model
1150
1151
1152
+def paddleocr_vl_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
1153
+ """Patches the HfRunner to fix create_causal_mask API mismatch.
1154
+
1155
+ The PaddleOCR-VL HF model passes `inputs_embeds` to create_causal_mask,
1156
+ but transformers renamed this parameter to `input_embeds`.
1157
+ """
1158
+ import sys
1159
+
1160
+ model_module = sys.modules.get(type(hf_model.model.model).__module__)
1161
+ if model_module is None:
1162
+ return hf_model
1163
+
1164
+ original_create_causal_mask = getattr(model_module, "create_causal_mask", None)
1165
+ if original_create_causal_mask is None:
1166
+ return hf_model
1167
+
1168
+ def patched_create_causal_mask(*args, **kwargs):
1169
+ if "inputs_embeds" in kwargs:
1170
+ kwargs["input_embeds"] = kwargs.pop("inputs_embeds")
1171
+ return original_create_causal_mask(*args, **kwargs)
1172
+
1173
+ model_module.create_causal_mask = patched_create_causal_mask # type: ignore[attr-defined]
1174
+ return hf_model
1175
+
1176
+
1177
def qwen2_5_omni_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
1178
"""Patches and returns an instance of the HfRunner for Qwen2.5-Omni."""
1179
thinker = hf_model.model.thinker