[CI/Build] Add Model Tests for Qwen2-VL (#9846)

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: DarkLight1337 <tlleungac@connect.ust.hk>

tests/models/decoder_only/vision_language/vlm_utils/model_utils.py

@@ -56,6 +56,17 @@ def qwen_vllm_to_hf_output(
     return output_ids, hf_output_str, out_logprobs
 
 
+def qwen2_vllm_to_hf_output(
+        vllm_output: RunnerOutput,
+        model: str) -> Tuple[List[int], str, Optional[SampleLogprobs]]:
+    """Sanitize vllm output [qwen2 models] to be comparable with hf output."""
+    output_ids, output_str, out_logprobs = vllm_output
+
+    hf_output_str = output_str + "<|im_end|>"
+
+    return output_ids, hf_output_str, out_logprobs
+
+
 def llava_image_vllm_to_hf_output(vllm_output: RunnerOutput,
                                   model: str) -> RunnerOutput:
     config = AutoConfig.from_pretrained(model)

tests/models/decoder_only/vision_language/vlm_utils/runners.py

@@ -29,6 +29,7 @@ def run_single_image_test(*, tmp_path: PosixPath, model_test_info: VLMTestInfo,
         num_logprobs=test_case.num_logprobs,
         limit_mm_per_prompt={"image": 1},
         distributed_executor_backend=test_case.distributed_executor_backend,
+        runner_mm_key="images",
         **model_test_info.get_non_parametrized_runner_kwargs())
 
 
@@ -51,6 +52,7 @@ def run_multi_image_test(*, tmp_path: PosixPath, model_test_info: VLMTestInfo,
         num_logprobs=test_case.num_logprobs,
         limit_mm_per_prompt={"image": len(image_assets)},
         distributed_executor_backend=test_case.distributed_executor_backend,
+        runner_mm_key="images",
         **model_test_info.get_non_parametrized_runner_kwargs())
 
 
@@ -74,6 +76,7 @@ def run_embedding_test(*, model_test_info: VLMTestInfo,
         limit_mm_per_prompt={"image": 1},
         vllm_embeddings=vllm_embeddings,
         distributed_executor_backend=test_case.distributed_executor_backend,
+        runner_mm_key="images",
         **model_test_info.get_non_parametrized_runner_kwargs())
 
 
@@ -101,6 +104,7 @@ def run_video_test(
         num_logprobs=test_case.num_logprobs,
         limit_mm_per_prompt={"video": len(video_assets)},
         distributed_executor_backend=test_case.distributed_executor_backend,
+        runner_mm_key="videos",
         **model_test_info.get_non_parametrized_runner_kwargs())
 
 
@@ -115,7 +119,11 @@ def run_custom_inputs_test(*, model_test_info: VLMTestInfo,
 
     inputs = test_case.custom_test_opts.inputs
     limit_mm_per_prompt = test_case.custom_test_opts.limit_mm_per_prompt
-    assert inputs is not None and limit_mm_per_prompt is not None
+    runner_mm_key = test_case.custom_test_opts.runner_mm_key
+    # Inputs, limit_mm_per_prompt, and runner_mm_key should all be set
+    assert inputs is not None
+    assert limit_mm_per_prompt is not None
+    assert runner_mm_key is not None
 
     core.run_test(
         hf_runner=hf_runner,
@@ -127,4 +135,5 @@ def run_custom_inputs_test(*, model_test_info: VLMTestInfo,
         num_logprobs=test_case.num_logprobs,
         limit_mm_per_prompt=limit_mm_per_prompt,
         distributed_executor_backend=test_case.distributed_executor_backend,
+        runner_mm_key=runner_mm_key,
         **model_test_info.get_non_parametrized_runner_kwargs())

tests/models/decoder_only/vision_language/vlm_utils/types.py

@@ -52,6 +52,8 @@ class SizeType(Enum):
 class CustomTestOptions(NamedTuple):
     inputs: List[Tuple[List[str], List[Union[List[Image], Image]]]]
     limit_mm_per_prompt: Dict[str, int]
+    # kwarg to pass multimodal data in as to vllm/hf runner instances.
+    runner_mm_key: str = "images"
 
 
 class ImageSizeWrapper(NamedTuple):
@@ -141,9 +143,6 @@ class VLMTestInfo(NamedTuple):
         Callable[[PosixPath, str, Union[List[ImageAsset], _ImageAssets]],
                  str]] = None  # noqa: E501
 
-    # kwarg to pass multimodal data in as to vllm/hf runner instances
-    runner_mm_key: str = "images"
-
     # Allows configuring a test to run with custom inputs
     custom_test_opts: Optional[List[CustomTestOptions]] = None
 
@@ -168,7 +167,6 @@ class VLMTestInfo(NamedTuple):
             "get_stop_token_ids": self.get_stop_token_ids,
             "model_kwargs": self.model_kwargs,
             "patch_hf_runner": self.patch_hf_runner,
-            "runner_mm_key": self.runner_mm_key,
         }