[CI/Build] Add text-only test for Qwen models (#7475)

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>

.buildkite/run-cpu-test.sh

@@ -22,7 +22,7 @@ docker exec cpu-test-avx2 bash -c "python3 examples/offline_inference.py"
 
 # Run basic model test
 docker exec cpu-test bash -c "
-  pip install pytest
+  pip install pytest matplotlib einops transformers_stream_generator
   pytest -v -s tests/models -m \"not vlm\" --ignore=tests/models/test_embedding.py --ignore=tests/models/test_oot_registration.py --ignore=tests/models/test_registry.py --ignore=tests/models/test_jamba.py --ignore=tests/models/test_danube3_4b.py" # Mamba and Danube3-4B on CPU is not supported
 
 # online inference

requirements-test.txt

@@ -11,7 +11,7 @@ pytest-shard
 
 # testing utils
 awscli
-einops # required for MPT
+einops # required for MPT and qwen-vl
 httpx
 peft
 requests
@@ -19,6 +19,8 @@ ray
 sentence-transformers # required for embedding
 compressed-tensors==0.4.0 # required for compressed-tensors
 timm # required for internvl test
+transformers_stream_generator # required for qwen-vl test
+matplotlib # required for qwen-vl test
 
 # TODO: Add this after fully implementing llava(mantis)
 # git+https://github.com/TIGER-AI-Lab/Mantis.git # required for llava(mantis) test

tests/models/test_qwen.py

+from typing import Type
+
+import pytest
+
+from ..conftest import HfRunner, VllmRunner
+from .utils import check_logprobs_close
+
+models = ["qwen/qwen-vl"]
+
+
+@pytest.mark.parametrize("dtype", ["half"])
+@pytest.mark.parametrize("max_tokens", [32])
+@pytest.mark.parametrize("num_logprobs", [5])
+@pytest.mark.parametrize("model", models)
+def test_text_only_qwen_model(
+    hf_runner: Type[HfRunner],
+    vllm_runner: Type[VllmRunner],
+    example_prompts,
+    model: str,
+    *,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+):
+    # This test checks language inputs only, since the visual component
+    # for qwen-vl is still unsupported in VLLM. In the near-future, the
+    # implementation and this test will be extended to consider
+    # visual inputs as well.
+    with hf_runner(model, dtype=dtype, is_vision_model=False) as hf_model:
+        hf_outputs = hf_model.generate_greedy_logprobs_limit(
+            example_prompts,
+            max_tokens,
+            num_logprobs=num_logprobs,
+        )
+
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy_logprobs(
+            example_prompts,
+            max_tokens,
+            num_logprobs=num_logprobs,
+        )
+
+    check_logprobs_close(
+        outputs_0_lst=hf_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )