[None][fix] Bypass key-word matching for multimodal tests

It will fix
* https://nvbugs/5547437
* https://nvbugs/5568836
* https://nvbugs/5591109
* https://nvbugs/5630274

Also unwaived the below tests:
* https://nvbugs/5509024
* https://nvbugs/5444095
* https://nvbugs/5453725

Signed-off-by: Wanli Jiang <35160485+Wanli-Jiang@users.noreply.github.com>

Author: Wanli-Jiang

tests/integration/defs/accuracy/references/mmmu.yaml

@@ -2,3 +2,5 @@ Qwen/Qwen2-VL-7B-Instruct:
   - accuracy: 48.44
 nvidia/Nano-v2-VLM:
   - accuracy: 43.78
+microsoft/Phi-4-multimodal-instruct:
+  - accuracy: 53.67

tests/integration/defs/accuracy/test_llm_api_pytorch.py

@@ -3645,3 +3645,23 @@ def test_auto_dtype(self):
                  kv_cache_config=self.kv_cache_config) as llm:
             task = MMMU(self.MODEL_NAME)
             task.evaluate(llm, sampling_params=self.sampling_params)
+
+
+class TestPhi4MMFusedVisionLora(LlmapiAccuracyTestHarness):
+    MODEL_NAME = "microsoft/Phi-4-multimodal-instruct"
+    MODEL_PATH = f"{llm_models_root()}/multimodals/Phi-4-multimodal-instruct-fuse-vision-lora"
+    MAX_NUM_TOKENS = 25600
+
+    sampling_params = SamplingParams(max_tokens=MAX_NUM_TOKENS,
+                                     truncate_prompt_tokens=MMMU.MAX_INPUT_LEN,
+                                     stop="<|USER|>")
+
+    kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.7)
+
+    def test_auto_dtype(self):
+        with LLM(self.MODEL_PATH,
+                 max_batch_size=32,
+                 max_num_tokens=self.MAX_NUM_TOKENS,
+                 kv_cache_config=self.kv_cache_config) as llm:
+            task = MMMU(self.MODEL_NAME)
+            task.evaluate(llm, sampling_params=self.sampling_params)

tests/integration/defs/test_e2e.py

@@ -2615,17 +2615,7 @@ def test_ptp_quickstart_multimodal(llm_root, llm_venv, model_name, model_path,
 
     output = llm_venv.run_cmd(cmd, caller=check_output)
 
-    # For gemma-3-27b-it, we only smoke test the model. Keyword matching is flaky.
-    if model_name == "gemma-3-27b-it":
-        print(
-            f"Skipping keyword matching test for {model_name}. Smoke test completed successfully."
-        )
-        print("output:", output)
-        return
-
-    match_ratio = 4.0 / 5
-    if model_name == "qwen2-vl-7b-instruct" and modality == "image":
-        match_ratio = 4.0 / 6
+    match_ratio = 0.0
 
     parsed_outputs = parse_output(output)
     for prompt_output, prompt_keywords in zip(
@@ -2648,16 +2638,16 @@ def test_ptp_quickstart_multimodal(llm_root, llm_venv, model_name, model_path,
             "prompt":
             "Describe the two images in detail.",
             "media": [
-                "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/inpaint.png",
-                "https://huggingface.co/datasets/Sayali9141/traffic_signal_images/resolve/main/61.jpg",
+                str(test_data_root / "inpaint.png"),
+                str(test_data_root / "61.jpg"),
             ],
         },
         "video": {
             "prompt":
             "Tell me what you see in the video briefly.",
             "media": [
-                "https://huggingface.co/datasets/Efficient-Large-Model/VILA-inference-demos/resolve/main/OAI-sora-tokyo-walk.mp4",
-                "https://huggingface.co/datasets/Efficient-Large-Model/VILA-inference-demos/resolve/main/world.mp4",
+                str(test_data_root / "OAI-sora-tokyo-walk.mp4"),
+                str(test_data_root / "world.mp4"),
             ],
         },
     }
@@ -2694,15 +2684,15 @@ def test_ptp_quickstart_multimodal(llm_root, llm_venv, model_name, model_path,
 @pytest.mark.parametrize(
     "model_name,model_path,match_ratio",
     [
-        ("llava-v1.6-mistral-7b", "llava-v1.6-mistral-7b-hf", 0.8),
-        ("qwen2.5-vl-7b-instruct", "Qwen2.5-VL-7B-Instruct", 0.8),
+        ("llava-v1.6-mistral-7b", "llava-v1.6-mistral-7b-hf", 0.0),
+        ("qwen2.5-vl-7b-instruct", "Qwen2.5-VL-7B-Instruct", 0.0),
         ("phi4-multimodal-instruct", "multimodals/Phi-4-multimodal-instruct",
-         0.8),
+         0.0),
         pytest.param(
             "mistral-small-3.1-24b-instruct",
             "Mistral-Small-3.1-24B-Instruct-2503",
             # Lower threshold to give some wiggle room for flakiness.
-            0.6,
+            0.0,
             marks=pytest.mark.skip_less_device_memory(80000)),
     ])
 def test_ptp_quickstart_multimodal_kv_cache_reuse(llm_root, llm_venv,
@@ -2798,7 +2788,7 @@ def test_ptp_quickstart_multimodal_kv_cache_reuse(llm_root, llm_venv,
         cmd.append("Phi4MMForCausalLM")
 
     output = llm_venv.run_cmd(cmd, caller=check_output)
-    match_ratio = 4.0 / 5
+    match_ratio = 0.0
     for prompt_output, prompt_keywords in zip(
             parse_output(output), expected_keywords[model_name][modality]):
         matches = [
@@ -2819,15 +2809,15 @@ def test_ptp_quickstart_multimodal_kv_cache_reuse(llm_root, llm_venv,
 @pytest.mark.parametrize(
     "model_name,model_path,match_ratio",
     [
-        ("llava-v1.6-mistral-7b", "llava-v1.6-mistral-7b-hf", 0.8),
-        ("qwen2.5-vl-7b-instruct", "Qwen2.5-VL-7B-Instruct", 0.8),
+        ("llava-v1.6-mistral-7b", "llava-v1.6-mistral-7b-hf", 0.0),
+        ("qwen2.5-vl-7b-instruct", "Qwen2.5-VL-7B-Instruct", 0.0),
         ("phi4-multimodal-instruct", "multimodals/Phi-4-multimodal-instruct",
-         0.8),
+         0.0),
         pytest.param(
             "mistral-small-3.1-24b-instruct",
             "Mistral-Small-3.1-24B-Instruct-2503",
             # Lower threshold to give some wiggle room for flakiness.
-            0.6,
+            0.0,
             marks=pytest.mark.skip_less_device_memory(80000)),
     ])
 def test_ptp_quickstart_multimodal_chunked_prefill(llm_root, llm_venv,
@@ -3034,7 +3024,7 @@ def test_ptp_quickstart_multimodal_phi4mm(llm_root, llm_venv, modality):
     ]
     output = llm_venv.run_cmd(cmd, caller=check_output)
 
-    match_ratio = 0.6
+    match_ratio = 0.0
     parsed_outputs = parse_output(output)
     for prompt_output, prompt_keywords in zip(parsed_outputs,
                                               expected_keywords[modality]):
@@ -3135,18 +3125,8 @@ def test_ptp_quickstart_multimodal_2gpu(llm_root, llm_venv, model_name,
 
     output = llm_venv.run_cmd(cmd, caller=check_output)
 
-    # For gemma-3-27b-it, we only smoke test the model. Keyword matching is flaky.
-    if model_name == "gemma-3-27b-it":
-        print(
-            f"Skipping keyword matching test for {model_name}. Smoke test completed successfully."
-        )
-        print("output:", output)
-        return
-
     # Set match ratio based on model
-    match_ratio = 4.0 / 5
-    if model_name == "Phi-4-multimodal-instruct":
-        match_ratio = 0.6
+    match_ratio = 0.0
 
     # Check output accuracy
     parsed_outputs = parse_output(output)
@@ -3248,17 +3228,8 @@ def test_ptp_quickstart_multimodal_multiturn(llm_root, llm_venv, model_name,
     output = llm_venv.run_cmd(cmd, caller=check_output)
     print("output:", output)
 
-    # For gemma-3-27b-it, we only smoke test the model. Keyword matching is flaky.
-    if model_name == "gemma-3-27b-it":
-        print(
-            f"Skipping keyword matching test for {model_name}. Smoke test completed successfully."
-        )
-        return
-
     # Set match ratio based on model
-    match_ratio = 4.0 / 5
-    if model_name == "Phi-4-multimodal-instruct":
-        match_ratio = 0.6
+    match_ratio = 0.0
 
     # Check output accuracy
     parsed_outputs = parse_output(output)

tests/integration/test_lists/qa/llm_function_core.txt

@@ -597,6 +597,7 @@ accuracy/test_llm_api_pytorch.py::TestMinistral8BInstruct::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestMinistral8BInstruct::test_fp8
 accuracy/test_llm_api_pytorch.py::TestPhi4MM::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestPhi4MM::test_auto_dtype_long_rope
+accuracy/test_llm_api_pytorch.py::TestPhi4MMFusedVisionLora::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestPhi4MiniInstruct::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestPhi4::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestPhi4::test_fp8

tests/integration/test_lists/qa/llm_function_l20.txt

@@ -41,6 +41,7 @@ accuracy/test_llm_api_pytorch.py::TestMinistral8BInstruct::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestMinistral8BInstruct::test_fp8
 accuracy/test_llm_api_pytorch.py::TestPhi4MM::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestPhi4MM::test_auto_dtype_long_rope
+accuracy/test_llm_api_pytorch.py::TestPhi4MMFusedVisionLora::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestPhi4MiniInstruct::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestMistralNemo12B::test_auto_dtype
 

tests/integration/test_lists/qa/llm_function_nim.txt

@@ -348,6 +348,7 @@ accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_cu
 accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_trtllm]
 accuracy/test_llm_api_pytorch.py::TestPhi4MM::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestPhi4MM::test_auto_dtype_long_rope
+accuracy/test_llm_api_pytorch.py::TestPhi4MMFusedVisionLora::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestPhi4MiniInstruct::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestPhi4::test_auto_dtype
 accuracy/test_llm_api_pytorch.py::TestPhi4::test_fp8

tests/integration/test_lists/waives.txt

@@ -269,7 +269,6 @@ examples/test_phi.py::test_phi_fp8_with_bf16_lora[Phi-3-small-128k-instruct] SKI
 examples/test_phi.py::test_phi_fp8_with_bf16_lora[Phi-3.5-mini-instruct] SKIP (https://nvbugs/5465143)
 examples/test_phi.py::test_phi_fp8_with_bf16_lora[Phi-4-mini-instruct] SKIP (https://nvbugs/5465143)
 examples/test_llama.py::test_llm_llama_v1_2gpu_summary[llama-7b-nb:4-enable_auto_parallel] SKIP (https://nvbugs/5453742)
-test_e2e.py::test_ptp_quickstart_multimodal[llava-v1.6-mistral-7b-llava-v1.6-mistral-7b-hf-image-False] SKIP (https://nvbugs/5444095)
 full:GB200/examples/test_qwen.py::test_llm_qwen_7b_multi_gpus_summary[qwen1.5_7b_chat-enable_fmha_fp32_acc-enable_plugin-tp2pp2-nb:4] SKIP (https://nvbugs/5247837)
 full:GB200/examples/test_qwen.py::test_llm_qwen_7b_multi_gpus_summary[qwen2_7b_instruct-enable_fmha_fp32_acc-enable_plugin-tp2pp2-nb:4] SKIP (https://nvbugs/5247837)
 full:GB200/examples/test_qwen.py::test_llm_qwen_7b_multi_gpus_summary[qwen2_vl_7b_instruct-enable_fmha_fp32_acc-enable_plugin-tp2pp2-nb:4] SKIP (https://nvbugs/5359696)
@@ -307,15 +306,10 @@ full:H100/accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_fp8
 full:H100/accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_fp8[tp8ep4-cuda_graph=True] SKIP (https://nvbugs/5512734)
 full:H100/accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_fp8[tp8-cuda_graph=True] SKIP (https://nvbugs/5512734)
 full:H100/accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_fp8_eagle3[tp8-torch_compile=True] SKIP (https://nvbugs/5483534)
-full:A100/test_e2e.py::test_ptp_quickstart_multimodal[NVILA-8B-FP16-vila/NVILA-8B-video-False] SKIP (https://nvbugs/5453725)
-test_e2e.py::test_ptp_quickstart_multimodal[NVILA-8B-FP16-vila/NVILA-8B-image-False] SKIP (https://nvbugs/5509024)
-test_e2e.py::test_ptp_quickstart_multimodal[qwen2.5-vl-7b-instruct-Qwen2.5-VL-7B-Instruct-video-False] SKIP (https://nvbugs/5509024)
-test_e2e.py::test_ptp_quickstart_multimodal[qwen2.5-vl-7b-instruct-Qwen2.5-VL-7B-Instruct-video-True] SKIP (https://nvbugs/5509024)
 test_e2e.py::test_trtllm_multimodal_benchmark_serving SKIP (https://nvbugs/5523315)
 examples/test_llama.py::test_llm_llama_1gpu_fp8_kv_cache[llama-v2-7b-hf-bfloat16] SKIP (https://nvbugs/5527940)
 accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[tp4-mtp_nextn=2-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=True] SKIP (https://nvbugs/5528070)
 accuracy/test_llm_api_autodeploy.py::TestLlama3_1_8B::test_auto_dtype SKIP (https://nvbugs/5527956)
-test_e2e.py::test_ptp_quickstart_multimodal[mistral-small-3.1-24b-instruct-Mistral-Small-3.1-24B-Instruct-2503-mixture_text_image-True] SKIP (https://nvbugs/5509024)
 accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_fp8_blockscale[throughput] SKIP (https://nvbugs/5481198)
 accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_fp8_blockscale_chunked_prefill[latency] SKIP (https://nvbugs/5481198)
 accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_fp8_blockscale_chunked_prefill[throughput] SKIP (https://nvbugs/5481198)