[Test][VLM] Add batch-invariance tests for Qwen3-VL (#53531)

Signed-off-by: machero <sr1123640465@126.com>
Co-authored-by: Claude <noreply@anthropic.com>

.buildkite/test_areas/misc.yaml

@@ -435,6 +435,7 @@ steps:
     - pip install pytest-timeout pytest-forked
     - pytest -v -s v1/determinism/test_batch_invariance.py
     - pytest -v -s v1/determinism/test_rms_norm_batch_invariant.py
+    - pytest -v -s v1/determinism/test_batch_invariance_vlm.py
     - VLLM_TEST_MODEL=deepseek-ai/DeepSeek-V2-Lite-Chat pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle -k TRITON_MLA
     - VLLM_TEST_MODEL=Qwen/Qwen3-30B-A3B-Thinking-2507-FP8 pytest -v -s v1/determinism/test_batch_invariance.py::test_v1_generation_is_deterministic_across_batch_sizes_with_needle -k FLASH_ATTN
     - pytest -v -s v1/determinism/test_nvfp4_batch_invariant.py

docs/features/batch_invariance.md

@@ -102,6 +102,7 @@ Batch invariance has been tested and verified on the following models:
 
 - **DeepSeek series**: `deepseek-ai/DeepSeek-V3`, `deepseek-ai/DeepSeek-V3-0324`, `deepseek-ai/DeepSeek-R1`, `deepseek-ai/DeepSeek-V3.1`
 - **Qwen3 (Dense)**: `Qwen/Qwen3-1.7B`, `Qwen/Qwen3-8B`, `Qwen/Qwen3-4B-AWQ`, `Qwen/Qwen3-8B-AWQ`
+- **Qwen3-VL (Vision-Language)**: `Qwen/Qwen3-VL-2B-Instruct`, `Qwen/Qwen3-VL-4B-Instruct` (single image and video inputs)
 - **Qwen3 (MoE)**: `Qwen/Qwen3-30B-A3B`, `Qwen/Qwen3-Next-80B-A3B-Instruct`, `Qwen/Qwen3-30B-A3B-Thinking-2507-FP8`
 - **Qwen2.5**: `Qwen/Qwen2.5-0.5B-Instruct`, `Qwen/Qwen2.5-1.5B-Instruct`, `Qwen/Qwen2.5-3B-Instruct`, `Qwen/Qwen2.5-7B-Instruct`, `Qwen/Qwen2.5-14B-Instruct`, `Qwen/Qwen2.5-32B-Instruct`
 - **Llama 3**: Llama3.1 and 3.2 series, `meta-llama/Llama-3.2-3B-Instruct` for example

tests/v1/determinism/test_batch_invariance_vlm.py

@@ -0,0 +1,103 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Batch-invariance regression tests for vision-language models (VLMs)."""
+
+import contextlib
+import os
+
+import numpy as np
+import pytest
+import torch
+from PIL import Image
+from utils import _extract_step_logprobs, skip_if_not_cuda
+
+from vllm import LLM, SamplingParams
+
+VLM_TEST_MODEL = os.getenv("VLLM_VLM_TEST_MODEL", "Qwen/Qwen3-VL-2B-Instruct")
+
+IMAGE_PROMPT = (
+    "<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>"
+    "Describe this image.<|im_end|><|im_start|>assistant\n"
+)
+VIDEO_PROMPT = (
+    "<|im_start|>user\n<|vision_start|><|video_pad|><|vision_end|>"
+    "Describe this video.<|im_end|><|im_start|>assistant\n"
+)
+
+
+def _make_image(seed: int, size: int) -> Image.Image:
+    rng = np.random.default_rng(seed)
+    return Image.fromarray(rng.integers(0, 255, (size, size, 3), dtype=np.uint8))
+
+
+def _make_inputs(input_type: str, num_reqs: int) -> list[dict]:
+    inputs = []
+    for i in range(num_reqs):
+        if input_type == "image":
+            prompt = IMAGE_PROMPT
+            mm_data = {"image": _make_image(1000 + i, size=256)}
+        else:
+            frames = [_make_image(1000 + i + f, size=128) for f in range(4)]
+            prompt = VIDEO_PROMPT
+            mm_data = {
+                "video": (
+                    frames,
+                    {
+                        "total_num_frames": len(frames),
+                        "fps": 1.0,
+                        "frames_indices": list(range(len(frames))),
+                        "do_sample_frames": False,
+                    },
+                )
+            }
+        inputs.append({"prompt": prompt, "multi_modal_data": mm_data})
+    return inputs
+
+
+def _assert_batch_invariant(llm: LLM, inputs: list[dict], sampling) -> None:
+    bs1 = [
+        _extract_step_logprobs(llm.generate([inp], sampling, use_tqdm=False)[0])
+        for inp in inputs
+    ]
+    bsN = [_extract_step_logprobs(out) for out in llm.generate(inputs, sampling)]
+
+    failures = []
+    for i, ((lp1, t1), (lpN, tN)) in enumerate(zip(bs1, bsN)):
+        if lp1 is None or lpN is None:
+            pytest.skip(
+                "Logits are not available on RequestOutput; "
+                "enable logprobs return to run this test."
+            )
+        if t1 != tN:
+            failures.append(f"req {i}: token mismatch bs1={t1} bsN={tN}")
+        elif not torch.equal(lp1, lpN):
+            d = (lp1 - lpN).abs().max().item()
+            failures.append(f"req {i}: logprob mismatch max_diff={d:.3e}")
+
+    if failures:
+        pytest.fail(
+            f"Batch invariance violated for {len(failures)}/{len(inputs)} "
+            f"requests:\n" + "\n".join(failures)
+        )
+
+
+@skip_if_not_cuda
+@pytest.mark.parametrize("input_type", ["image", "video"])
+@pytest.mark.parametrize("mm_encoder_attn_backend", ["FLASH_ATTN", "TORCH_SDPA"])
+def test_vlm_batch_invariance_bs1_vs_bsN(input_type: str, mm_encoder_attn_backend: str):
+    inputs = _make_inputs(input_type, num_reqs=4)
+    sampling = SamplingParams(temperature=0.0, max_tokens=8, seed=1234, logprobs=5)
+
+    llm = LLM(
+        model=VLM_TEST_MODEL,
+        dtype="bfloat16",
+        enforce_eager=True,
+        max_num_seqs=16,
+        gpu_memory_utilization=0.85,
+        mm_encoder_attn_backend=mm_encoder_attn_backend,
+    )
+    try:
+        _assert_batch_invariant(llm, inputs, sampling)
+    finally:
+        with contextlib.suppress(Exception):
+            llm.shutdown()