[https://nvbugs/5608723][fix] Use local data on multimodal tests and unwaive tests (#8673)

Signed-off-by: yechank <161688079+yechank-nvidia@users.noreply.github.com>

Author: yechank-nvidia

tests/integration/defs/test_e2e.py

@@ -1665,7 +1665,6 @@ def test_openai_lora(llm_root, llm_venv):
     llm_venv.run_cmd(["-m", "pytest", str(test_root / "_test_openai_lora.py")])
 
 
-@pytest.mark.skip(reason="https://nvbugs/5596377")
 def test_openai_chat_multimodal_example(llm_root, llm_venv):
     test_root = unittest_path() / "llmapi" / "apps"
     llm_venv.run_cmd(

tests/unittest/_torch/modeling/test_modeling_qwen2_5vl.py

@@ -5,13 +5,13 @@
 from pathlib import Path
 from typing import List
 
-import pytest
 import torch
 from _torch.helpers import create_mock_engine
 from parameterized import parameterized
 from transformers import AutoProcessor, AutoTokenizer, Qwen2_5_VLConfig
 from transformers import \
     Qwen2_5_VLForConditionalGeneration as HFQwen2_5_VLForConditionalLM
+from utils.llm_data import llm_models_root
 
 import tensorrt_llm
 from tensorrt_llm._torch.attention_backend.utils import get_attention_backend
@@ -28,16 +28,6 @@
 from tensorrt_llm.inputs.multimodal import MultimodalParams
 from tensorrt_llm.mapping import Mapping
 
-
-def llm_models_root() -> str:
-    '''return LLM_MODELS_ROOT path if it is set in env, assert when it's set but not a valid path
-    '''
-    DEFAULT_LLM_MODEL_ROOT = os.path.join("/scratch.trt_llm_data", "llm-models")
-    LLM_MODELS_ROOT = os.environ.get("LLM_MODELS_ROOT", DEFAULT_LLM_MODEL_ROOT)
-
-    return LLM_MODELS_ROOT
-
-
 QWEN2_5_VL_7B_CONFIG = {
     "architectures": ["Qwen2_5_VLForConditionalGeneration"],
     "attention_dropout": 0.0,
@@ -222,7 +212,8 @@ def get_hf_inputs(self, model, modality: str, device: torch.device,
             images = [input['multi_modal_data']['image'] for input in inputs]
         elif modality == "video":
             videos = [
-                input['multi_modal_data'][f'{modality}'] for input in inputs
+                input['multi_modal_data'][f'{modality}'][0].frames
+                for input in inputs
             ]
         elif modality == "text":
             # For text-only modality, no images or videos needed
@@ -240,7 +231,6 @@ def get_hf_inputs(self, model, modality: str, device: torch.device,
         ).to(device)
         return processor_inputs
 
-    @pytest.mark.skip(reason="https://nvbugs/5550722")
     def test_qwen2_5_vl_sanity(self):
 
         config_dict = deepcopy(QWEN2_5_VL_7B_CONFIG)
@@ -359,7 +349,6 @@ def test_qwen2_5_vl_sanity(self):
                  use_cuda_graph=False,
                  disable_fuse_rope=False),
     ])
-    @pytest.mark.skip(reason="https://nvbugs/5550722")
     @torch.no_grad()
     def test_qwen2_5_vl_allclose_to_hf(self, scenario: Scenario) -> None:
         """

tests/unittest/_torch/multimodal/test_mm_encoder_standalone.py

@@ -177,7 +177,7 @@ def test_multi_request_batch_chat(model_key, multimodal_model_config):
             f"Skipping test for {model_key} - only testing llava-v1.6-mistral-7b-hf for now"
         )
 
-    encoder_model_dir = multimodal_model_config['hf_model_dir']
+    encoder_model_dir = multimodal_model_config['model_dir']
 
     max_tokens = 64
     free_gpu_memory_fraction = 0.6