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 3, 19:00 UTC
[CI] Re-enable Nemotron Parse parity test and switch testing to nemotron-parse v1.2 (#42498)
Signed-off-by: <mwawrzos@nvidia.com>
Marek Wawrzos committed
May 13, 2026 at 15:05 UTC
67671692ace71347797b0372376b0f194c924ac8
3 files changed
+12
-9
tests/conftest.py
+6
-1
@@ -785,10 +785,15 @@ class HfRunner:
785
audios: PromptAudioInput | None = None,
786
videos: PromptVideoInput | None = None,
787
use_cache: bool = True,
788
+ tokenization_kwargs: dict[str, Any] | None = None,
789
**kwargs: Any,
790
) -> list[TokensTextLogprobs]:
791
all_inputs = self.get_inputs(
791
- prompts, images=images, videos=videos, audios=audios
792
+ prompts,
793
+ images=images,
794
+ videos=videos,
795
+ audios=audios,
796
+ tokenization_kwargs=tokenization_kwargs,
797
)
798
799
all_logprobs: list[list[dict[int, float]]] = []
tests/models/multimodal/generation/test_nemotron_parse.py
+5
-7
@@ -14,7 +14,9 @@ from vllm.tokenizers import TokenizerLike
14
from ....conftest import HfRunner, PromptImageInput, VllmRunner
15
16
IMAGE = ImageAsset("paper-11").pil_image_ext(ext="png").convert("RGB")
17
-PROMPT = "</s><s><predict_bbox><predict_classes><output_markdown>"
17
+PROMPT = (
18
+ "</s><s><predict_bbox><predict_classes><output_markdown><predict_no_text_in_pic>"
19
+)
20
21
22
class DummyLogprobs(dict[int, Logprob]):
@@ -85,7 +87,7 @@ def run_test(
87
max_tokens,
88
num_logprobs=num_logprobs,
89
images=images,
88
- use_cache=False, # HF Nemotron Parse crashes here without this
90
+ tokenization_kwargs={"add_special_tokens": False},
91
)
92
for prompts, images in inputs
93
]
@@ -103,11 +105,7 @@ def run_test(
105
)
106
107
106
-@pytest.mark.skip(
107
- reason="Model's custom MBart decoder has head count mismatch with "
108
- "transformers v5's GQA-aware cross-attention (8 vs 16 heads)"
109
-)
110
-@pytest.mark.parametrize("model", ["nvidia/NVIDIA-Nemotron-Parse-v1.1"])
108
+@pytest.mark.parametrize("model", ["nvidia/NVIDIA-Nemotron-Parse-v1.2"])
109
@pytest.mark.parametrize("dtype", ["bfloat16"])
110
@pytest.mark.parametrize("num_logprobs", [5])
111
def test_models(
tests/models/registry.py
+1
-1
@@ -1379,7 +1379,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
1379
"CohereLabs/cohere-transcribe-03-2026", trust_remote_code=True
1380
),
1381
"NemotronParseForConditionalGeneration": _HfExamplesInfo(
1382
- "nvidia/NVIDIA-Nemotron-Parse-v1.1", trust_remote_code=True
1382
+ "nvidia/NVIDIA-Nemotron-Parse-v1.2", trust_remote_code=True
1383
),
1384
"WhisperForConditionalGeneration": _HfExamplesInfo(
1385
"openai/whisper-large-v3-turbo",