From f99762eb254807335419cd04945359eaa5213713 Mon Sep 17 00:00:00 2001
From: Canlin Guo <canlinguosdu@gmail.com>
Date: Fri, 31 Oct 2025 15:42:47 +0800
Subject: [PATCH] [E2E][MM] Add e2e tests for InternVL model (#3796)

### What this PR does / why we need it?

As a validation for #3664, add end-to-end tests to monitor the InternVL
model and ensure its continuous proper operation. This PR is only for
single-card. So the models that have more parameters than 8B like 78B
are needed to test using multi-cards.


### Does this PR introduce _any_ user-facing change?

None.

### How was this patch tested?

`pytest -sv tests/e2e/singlecard/multi-modal/test_internvl.py`


- vLLM version: v0.11.0
- vLLM main:
https://github.com/vllm-project/vllm/commit/83f478bb19489b41e9d208b47b4bb5a95ac171ac

---------

Signed-off-by: gcanlin <canlinguosdu@gmail.com>
---
 .github/workflows/_e2e_test.yaml              |  1 +
 .../singlecard/multi-modal/test_internvl.py   | 97 +++++++++++++++++++
 2 files changed, 98 insertions(+)
 create mode 100644 tests/e2e/singlecard/multi-modal/test_internvl.py

diff --git a/.github/workflows/_e2e_test.yaml b/.github/workflows/_e2e_test.yaml
index 8d383cff..cb2582e9 100644
--- a/.github/workflows/_e2e_test.yaml
+++ b/.github/workflows/_e2e_test.yaml
@@ -103,6 +103,7 @@ jobs:
           pytest -sv tests/e2e/singlecard/test_quantization.py
           pytest -sv tests/e2e/singlecard/test_sampler.py
           pytest -sv tests/e2e/singlecard/test_vlm.py
+          pytest -sv tests/e2e/singlecard/multi-modal/test_internvl.py
 
           # ------------------------------------ v1 spec decode test ------------------------------------ #
           pytest -sv tests/e2e/singlecard/spec_decode_v1/test_v1_mtp_correctness.py
diff --git a/tests/e2e/singlecard/multi-modal/test_internvl.py b/tests/e2e/singlecard/multi-modal/test_internvl.py
new file mode 100644
index 00000000..7cf32e4d
--- /dev/null
+++ b/tests/e2e/singlecard/multi-modal/test_internvl.py
@@ -0,0 +1,97 @@
+#
+# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# This file is a part of the vllm-ascend project.
+#
+
+import os
+
+# Set spawn method before any torch/NPU imports to avoid fork issues
+os.environ.setdefault('VLLM_WORKER_MULTIPROC_METHOD', 'spawn')
+
+import pytest
+from vllm.assets.image import ImageAsset
+
+from tests.e2e.conftest import VllmRunner
+from tests.e2e.model_utils import check_outputs_equal
+from vllm_ascend.utils import vllm_version_is
+
+MODELS = [
+    "OpenGVLab/InternVL2-8B",
+    "OpenGVLab/InternVL2_5-8B",
+    "OpenGVLab/InternVL3-8B",
+    "OpenGVLab/InternVL3_5-8B",
+]
+
+# skip testing InternVL3-8B and InternVL3_5-8B on 0.11.0 due to https://github.com/vllm-project/vllm-ascend/issues/3925.
+if vllm_version_is("0.11.0"):
+    MODELS = [
+        "OpenGVLab/InternVL2-8B",
+        "OpenGVLab/InternVL2_5-8B",
+    ]
+
+
+@pytest.mark.parametrize("model", MODELS)
+def test_internvl_basic(model: str):
+    """Test basic InternVL2 inference with single image."""
+    # Load test image
+    image = ImageAsset("cherry_blossom").pil_image.convert("RGB")
+
+    # InternVL uses chat template format
+    # Format: <|im_start|>user\n<image>\nQUESTION<|im_end|>\n<|im_start|>assistant\n
+    questions = [
+        "What is the content of this image?",
+        "Describe this image in detail.",
+    ]
+
+    # Build prompts with InternVL2 chat template
+    prompts = [
+        f"<|im_start|>user\n<image>\n{q}<|im_end|>\n<|im_start|>assistant\n"
+        for q in questions
+    ]
+    images = [image] * len(prompts)
+
+    outputs = {}
+    for enforce_eager, mode in [(False, "eager"), (True, "graph")]:
+        with VllmRunner(
+                model,
+                max_model_len=8192,
+                limit_mm_per_prompt={"image": 4},
+                enforce_eager=enforce_eager,
+                dtype="bfloat16",
+        ) as vllm_model:
+            generated_outputs = vllm_model.generate_greedy(
+                prompts=prompts,
+                images=images,
+                max_tokens=128,
+            )
+
+            assert len(generated_outputs) == len(prompts), \
+                f"Expected {len(prompts)} outputs, got {len(generated_outputs)} in {mode} mode"
+
+            for i, (_, output_str) in enumerate(generated_outputs):
+                assert output_str, \
+                    f"{mode.capitalize()} mode output {i} should not be empty. Prompt: {prompts[i]}"
+                assert len(output_str.strip()) > 0, \
+                    f"{mode.capitalize()} mode Output {i} should have meaningful content"
+
+            outputs[mode] = generated_outputs
+
+    eager_outputs = outputs["eager"]
+    graph_outputs = outputs["graph"]
+
+    check_outputs_equal(outputs_0_lst=eager_outputs,
+                        outputs_1_lst=graph_outputs,
+                        name_0="eager mode",
+                        name_1="graph mode")