[Bug][ROCm]: `vision_embeddings` in transformers inaccurate without math SDP

[AndreasKaratzas]

==============================
        System Info
==============================
OS                           : Ubuntu 22.04.5 LTS (x86_64)
GCC version                  : (Ubuntu 11.4.0-1ubuntu1~22.04.2) 11.4.0
Clang version                : 20.0.0git (https://github.com/RadeonOpenCompute/llvm-project roc-7.1.0 25425 1b0eada6b0ee93e2e694c8c146d23fca90bc11c5)
CMake version                : version 3.31.10
Libc version                 : glibc-2.35

==============================
       PyTorch Info
==============================
PyTorch version              : 2.9.0a0+git1c57644
Is debug build               : False
CUDA used to build PyTorch   : N/A
ROCM used to build PyTorch   : 7.1.25424-4179531dcd

==============================
      Python Environment
==============================
Python version               : 3.12.12 (main, Oct 10 2025, 08:52:57) [GCC 11.4.0] (64-bit runtime)
Python platform              : Linux-5.15.0-161-generic-x86_64-with-glibc2.35

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : Could not collect
CUDA_MODULE_LOADING set to   :
GPU models and configuration :  (gfx942:sramecc+:xnack-)
Nvidia driver version        : Could not collect
cuDNN version                : Could not collect
HIP runtime version          : 7.1.25424
MIOpen runtime version       : 3.5.1
Is XNNPACK available         : True

==============================
          CPU Info
==============================
Architecture:                            x86_64
CPU op-mode(s):                          32-bit, 64-bit
Address sizes:                           52 bits physical, 57 bits virtual
Byte Order:                              Little Endian
CPU(s):                                  256
On-line CPU(s) list:                     0-255
Vendor ID:                               AuthenticAMD
Model name:                              AMD EPYC 9575F 64-Core Processor
CPU family:                              26
Model:                                   2
Thread(s) per core:                      2
Core(s) per socket:                      64
Socket(s):                               2
Stepping:                                1
Frequency boost:                         enabled
CPU max MHz:                             5008.0068
CPU min MHz:                             1500.0000
BogoMIPS:                                6600.02
Flags:                                   fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monito
r ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pst
ate ssbd mba ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_m
bm_local avx_vnni avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclm
ulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid bus_lock_detect movdiri movdir64b overflow_recov succor smca fsrm avx512_vp2intersect flush_l1d
Virtualization:                          AMD-V
L1d cache:                               6 MiB (128 instances)
L1i cache:                               4 MiB (128 instances)
L2 cache:                                128 MiB (128 instances)
L3 cache:                                512 MiB (16 instances)
NUMA node(s):                            2
NUMA node0 CPU(s):                       0-63,128-191
NUMA node1 CPU(s):                       64-127,192-255
Vulnerability Gather data sampling:      Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Not affected
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:                Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIBP always-on; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Not affected

==============================
Versions of relevant libraries
==============================
[pip3] conch-triton-kernels==1.2.1
[pip3] efficientnet_pytorch==0.7.1
[pip3] numpy==2.2.6
[pip3] open_clip_torch==2.32.0
[pip3] pytorch-lightning==2.6.0
[pip3] pyzmq==27.1.0
[pip3] segmentation_models_pytorch==0.4.0
[pip3] sentence-transformers==5.1.2
[pip3] terratorch==1.0.2
[pip3] torch==2.9.0a0+git1c57644
[pip3] torchaudio==2.9.0+eaa9e4e
[pip3] torchgeo==0.7.0
[pip3] torchmetrics==1.8.2
[pip3] torchvision==0.23.0a0+824e8c8
[pip3] transformers==4.57.3
[pip3] triton==3.4.0
[pip3] triton_kernels==1.0.0
[conda] Could not collect
==============================
         vLLM Info
==============================
ROCM Version                 : 7.1.25424-4179531dcd
vLLM Version                 : 0.11.2.dev599+ge8dc42f0a.d20251206 (git sha: e8dc42f0a, date: 20251206)
vLLM Build Flags:
  CUDA Archs: Not Set; ROCm: Disabled
GPU Topology:
  ============================ ROCm System Management Interface ============================
================================ Weight between two GPUs =================================
       GPU0         GPU1         GPU2         GPU3         GPU4         GPU5         GPU6         GPU7
GPU0   0            15           15           15           15           15           15           15
GPU1   15           0            15           15           15           15           15           15
GPU2   15           15           0            15           15           15           15           15
GPU3   15           15           15           0            15           15           15           15
GPU4   15           15           15           15           0            15           15           15
GPU5   15           15           15           15           15           0            15           15
GPU6   15           15           15           15           15           15           0            15
GPU7   15           15           15           15           15           15           15           0

================================= Hops between two GPUs ==================================
       GPU0         GPU1         GPU2         GPU3         GPU4         GPU5         GPU6         GPU7
GPU0   0            1            1            1            1            1            1            1
GPU1   1            0            1            1            1            1            1            1
GPU2   1            1            0            1            1            1            1            1
GPU3   1            1            1            0            1            1            1            1
GPU4   1            1            1            1            0            1            1            1
GPU5   1            1            1            1            1            0            1            1
GPU6   1            1            1            1            1            1            0            1
GPU7   1            1            1            1            1            1            1            0

=============================== Link Type between two GPUs ===============================
       GPU0         GPU1         GPU2         GPU3         GPU4         GPU5         GPU6         GPU7
GPU0   0            XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         XGMI
GPU1   XGMI         0            XGMI         XGMI         XGMI         XGMI         XGMI         XGMI
GPU2   XGMI         XGMI         0            XGMI         XGMI         XGMI         XGMI         XGMI
GPU3   XGMI         XGMI         XGMI         0            XGMI         XGMI         XGMI         XGMI
GPU4   XGMI         XGMI         XGMI         XGMI         0            XGMI         XGMI         XGMI
GPU5   XGMI         XGMI         XGMI         XGMI         XGMI         0            XGMI         XGMI
GPU6   XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         0            XGMI
GPU7   XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         XGMI         0

======================================= Numa Nodes =======================================
GPU[0]          : (Topology) Numa Node: 0
GPU[0]          : (Topology) Numa Affinity: 0
GPU[1]          : (Topology) Numa Node: 0
GPU[1]          : (Topology) Numa Affinity: 0
GPU[2]          : (Topology) Numa Node: 0
GPU[2]          : (Topology) Numa Affinity: 0
GPU[3]          : (Topology) Numa Node: 0
GPU[3]          : (Topology) Numa Affinity: 0
GPU[4]          : (Topology) Numa Node: 1
GPU[4]          : (Topology) Numa Affinity: 1
GPU[5]          : (Topology) Numa Node: 1
GPU[5]          : (Topology) Numa Affinity: 1
GPU[6]          : (Topology) Numa Node: 1
GPU[6]          : (Topology) Numa Affinity: 1
GPU[7]          : (Topology) Numa Node: 1
GPU[7]          : (Topology) Numa Affinity: 1
================================== End of ROCm SMI Log ===================================

==============================
     Environment Variables
==============================
PYTORCH_ROCM_ARCH=gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151
LD_LIBRARY_PATH=/opt/rocm/lib:/usr/local/lib:
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1

🐛 Describe the bug

On ROCm platforms, multimodal models using the vLLM Transformers backend produce incorrect/garbage outputs due to accuracy issues with the flash_sdp and mem_efficient_sdp backends in PyTorch's scaled_dot_product_attention when used in vision encoders.

Affected Components

• File: vllm/model_executor/models/transformers/multimodal.py
• Method: MultiModalMixin.embed_multimodal()
• Affected models: Multimodal models using the Transformers backend (e.g., Qwen2.5-VL, potentially others)

Root Cause

When self.model.get_image_features() is called, the vision encoder (e.g., Qwen2_5_VLVisionAttention in HuggingFace Transformers) uses torch.nn.functional.scaled_dot_product_attention. On ROCm, this defaults to flash_sdp or mem_efficient_sdp backends which have numerical accuracy issues, causing the vision encoder to produce incorrect embeddings.

The LLM text decoder is unaffected because vLLM sets _attn_implementation = "vllm" for the text config, but the vision encoder config is not modified and uses the default SDPA backends.

Reproduction

Test file: tests/models/multimodal/generation/test_qwen25_vl_transformers.py

import math
from collections import defaultdict
from pathlib import PosixPath

import pytest
from transformers import (
    AutoModelForImageTextToText,
)

from ....conftest import (
    HfRunner,
    ImageTestAssets,
    VllmRunner,
)
from ....utils import large_gpu_mark
from .vlm_utils import model_utils, runners
from .vlm_utils.case_filtering import get_parametrized_options
from .vlm_utils.types import (
    ExpandableVLMTestArgs,
    VLMTestInfo,
    VLMTestType,
)

COMMON_BROADCAST_SETTINGS = {
    "test_type": VLMTestType.IMAGE,
    "dtype": "half",
    "max_tokens": 5,
    "tensor_parallel_size": 2,
    "hf_model_kwargs": {"device_map": "auto"},
    "image_size_factors": [(0.25, 0.5, 1.0)],
    "distributed_executor_backend": (
        "ray",
        "mp",
    ),
}

VLM_TEST_SETTINGS = {
    "qwen2_5_vl-transformers": VLMTestInfo(
        models=["Qwen/Qwen2.5-VL-3B-Instruct"],
        test_type=VLMTestType.IMAGE,
        prompt_formatter=lambda img_prompt: f"<|im_start|>User\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n",  # noqa: E501
        img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
        max_model_len=4096,
        max_num_seqs=2,
        auto_cls=AutoModelForImageTextToText,
        vllm_output_post_proc=model_utils.qwen2_vllm_to_hf_output,
        image_size_factors=[(0.25, 0.2, 0.15)],
        vllm_runner_kwargs={
            "model_impl": "transformers",
        },
        marks=[large_gpu_mark(32)],
    ),
}


def _mark_splits(
    test_settings: dict[str, VLMTestInfo],
    *,
    num_groups: int,
) -> dict[str, VLMTestInfo]:
    name_by_test_info_id = {id(v): k for k, v in test_settings.items()}
    test_infos_by_model = defaultdict[str, list[VLMTestInfo]](list)

    for info in test_settings.values():
        for model in info.models:
            test_infos_by_model[model].append(info)

    models = sorted(test_infos_by_model.keys())
    split_size = math.ceil(len(models) / num_groups)

    new_test_settings = dict[str, VLMTestInfo]()

    for i in range(num_groups):
        models_in_group = models[i * split_size : (i + 1) * split_size]

        for model in models_in_group:
            for info in test_infos_by_model[model]:
                new_marks = (info.marks or []) + [pytest.mark.split(group=i)]
                new_info = info._replace(marks=new_marks)
                new_test_settings[name_by_test_info_id[id(info)]] = new_info

    missing_keys = test_settings.keys() - new_test_settings.keys()
    assert not missing_keys, f"Missing keys: {missing_keys}"

    return new_test_settings


VLM_TEST_SETTINGS = _mark_splits(VLM_TEST_SETTINGS, num_groups=2)

@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.IMAGE,
        create_new_process_for_each_test=False,
    ),
)
def test_single_image_models(
    tmp_path: PosixPath,
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_single_image_test(
        tmp_path=tmp_path,
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
        image_assets=image_assets,
    )


@pytest.mark.parametrize(
    "model_type,test_case",
    get_parametrized_options(
        VLM_TEST_SETTINGS,
        test_type=VLMTestType.MULTI_IMAGE,
        create_new_process_for_each_test=False,
    ),
)
def test_multi_image_models(
    tmp_path: PosixPath,
    model_type: str,
    test_case: ExpandableVLMTestArgs,
    hf_runner: type[HfRunner],
    vllm_runner: type[VllmRunner],
    image_assets: ImageTestAssets,
):
    model_test_info = VLM_TEST_SETTINGS[model_type]
    runners.run_multi_image_test(
        tmp_path=tmp_path,
        model_test_info=model_test_info,
        test_case=test_case,
        hf_runner=hf_runner,
        vllm_runner=vllm_runner,
        image_assets=image_assets,
    )

Temporary Fix

In embed_multimodal() under vllm/model_executor/models/transformers/multimodal.py, wrap the get_image_features() call with torch.nn.attention.sdpa_kernel to force the MATH backend on ROCm:

if pixel_values is not None:
    if current_platform.is_rocm():
        with torch.nn.attention.sdpa_kernel(
            backends=[torch.nn.attention.SDPBackend.MATH]
        ):
            vision_embeddings = self.model.get_image_features(
                pixel_values, **kwargs
            )
    else:
        vision_embeddings = self.model.get_image_features(pixel_values, **kwargs)

Notes

• The MATH backend is slower but numerically accurate
• Setting _attn_implementation = "eager" on vision_config causes OOM if max_num_encoder_tokens is large enough due to full attention matrix materialization
• The mm_processor_kwargs with reduced max_pixels is needed to avoid OOM with the MATH backend during profiling

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[github-actions]

CC @hongxiayang @tjtanaa @vllmellm for ROCm-related issue

[DarkLight1337]

cc @hmellor

[hmellor]

Thanks for the ping, next week I'll look into adding the attention interface vLLM uses for vision attention to ALL_ATTENTION_FUNCTIONS and setting the _attention_interface to use it in the vision config.