[Nano-Nemotron] Fix Nano-Nemotron precomputed multimodal embeddings (#56398)

Signed-off-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>
This commit is contained in:
Taneem Ibrahim
2026-09-12 09:07:10 -04:00
committed by GitHub
parent bde4feb061
commit a0914ab7d0
2 changed files with 23 additions and 4 deletions
@@ -5,6 +5,7 @@ from unittest.mock import patch
import numpy as np
import pytest
import torch
from vllm import envs
from vllm.model_executor.models.nano_nemotron_vl import (
@@ -14,6 +15,17 @@ from vllm.model_executor.models.nano_nemotron_vl import (
from vllm.multimodal.parse import MultiModalDataItems, VideoProcessorItems
@pytest.mark.parametrize("input_key", ["image_embeds", "video_embeds"])
def test_precomputed_multimodal_embeddings(input_key: str):
model = object.__new__(NemotronH_Nano_VL_V2)
embeds = torch.randn(2, 4, 8)
outputs = model.embed_multimodal(**{input_key: embeds})
assert len(outputs) == len(embeds)
assert all(torch.equal(output, embed) for output, embed in zip(outputs, embeds))
class _TextOnlyMultiModalConfig:
def get_limit_per_prompt(self, modality: str) -> int:
return 0