[Bug]: Critical Memory Leak in vLLM V1 Engine: 200+ GB RAM Usage from Image Inference

[oyerli]

Your current environment

The output of `python collect_env.py`

Collecting environment information...
PyTorch version: 2.6.0+cu124
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.5 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.22.1
Libc version: glibc-2.35

Python version: 3.10.12 (main, Feb  4 2025, 14:57:36) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.0-130-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.1.105
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA H100 80GB HBM3
Nvidia driver version: 550.144.03
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.0
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        46 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               128
On-line CPU(s) list:                  0-127
Vendor ID:                            GenuineIntel
Model name:                           Intel(R) Xeon(R) Platinum 8462Y+
CPU family:                           6
Model:                                143
Thread(s) per core:                   2
Core(s) per socket:                   32
Socket(s):                            2
Stepping:                             8
Frequency boost:                      enabled
CPU max MHz:                          2801.0000
CPU min MHz:                          800.0000
BogoMIPS:                             5600.00
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 invpcid_single intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization:                       VT-x
L1d cache:                            3 MiB (64 instances)
L1i cache:                            2 MiB (64 instances)
L2 cache:                             128 MiB (64 instances)
L3 cache:                             120 MiB (2 instances)
NUMA node(s):                         2
NUMA node0 CPU(s):                    0-31,64-95
NUMA node1 CPU(s):                    32-63,96-127
Vulnerability Gather data sampling:   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; RSB filling; PBRSB-eIBRS SW sequence; BHI BHI_DIS_S
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] nvidia-cublas-cu12==12.4.5.8
[pip3] nvidia-cuda-cupti-cu12==12.4.127
[pip3] nvidia-cuda-nvrtc-cu12==12.4.127
[pip3] nvidia-cuda-runtime-cu12==12.4.127
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.2.1.3
[pip3] nvidia-curand-cu12==10.3.5.147
[pip3] nvidia-cusolver-cu12==11.6.1.9
[pip3] nvidia-cusparse-cu12==12.3.1.170
[pip3] nvidia-cusparselt-cu12==0.6.2
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] pyzmq==26.2.1
[pip3] torch==2.6.0
[pip3] torchaudio==2.6.0
[pip3] torchvision==0.21.0
[pip3] transformers==4.50.0
[pip3] triton==3.2.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.8.1
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      0-31,64-95      0               N/A

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

NVIDIA_VISIBLE_DEVICES=GPU-c19a2d90-c6b2-d624-af66-28214621a956
NVIDIA_REQUIRE_CUDA=cuda>=12.1 brand=tesla,driver>=470,driver<471 brand=unknown,driver>=470,driver<471 brand=nvidia,driver>=470,driver<471 brand=nvidiartx,driver>=470,driver<471 brand=geforce,driver>=470,driver<471 brand=geforcertx,driver>=470,driver<471 brand=quadro,driver>=470,driver<471 brand=quadrortx,driver>=470,driver<471 brand=titan,driver>=470,driver<471 brand=titanrtx,driver>=470,driver<471 brand=tesla,driver>=525,driver<526 brand=unknown,driver>=525,driver<526 brand=nvidia,driver>=525,driver<526 brand=nvidiartx,driver>=525,driver<526 brand=geforce,driver>=525,driver<526 brand=geforcertx,driver>=525,driver<526 brand=quadro,driver>=525,driver<526 brand=quadrortx,driver>=525,driver<526 brand=titan,driver>=525,driver<526 brand=titanrtx,driver>=525,driver<526
NCCL_VERSION=2.17.1-1
NVIDIA_DRIVER_CAPABILITIES=all
NVIDIA_PRODUCT_NAME=CUDA
CUDA_VERSION=12.1.1
LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
NCCL_CUMEM_ENABLE=0
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

We are using vLLM in production and yesterday upgraded to version 0.8.1. After switching to the new V1 engine, we observed a significant RAM memory leak during image inference requests, even though the model runs on the GPU.

In our production environment, RAM usage ballooned to over 200 GB. Upon further investigation, we reproduced the issue using a minimal example (provided below). With V1 engine, each image inference request increases system memory usage, whereas V0 engine does not have this issue.

Summary:

1. Issue only occurs in V1 engine.
2. Memory leak is tied to image inference requests.
3. Behavior is not present in V0.

We’d greatly appreciate it if you could investigate and let us know if you need any additional information. We’re happy to assist in resolving this important issue as soon as possible.

vLLM serve command (using V1):

vllm serve mistralai/Pixtral-12B-2409 --tokenizer_mode mistral --limit_mm_per_prompt 'image=4' --disable-mm-preprocessor-cache

Initial memory usage (3209 MB):

Run image inference 100 times:

import requests
import json

url = "http://127.0.0.1:8000/v1/chat/completions"
headers = {"Content-Type": "application/json", "Authorization": "Bearer token"}

model = "mistralai/Pixtral-12B-2409"

image_url = "https://texascoffeeschool.com/wp-content/uploads/2021/10/DSC_0052-scaled.jpg"

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "Describe this image in a single sentence.",
            },
            {"type": "image_url", "image_url": {"url": image_url}},
        ],
    },
]

data = {"model": model, "messages": messages}

for _ in range(1,100):
    response = requests.post(url, headers=headers, data=json.dumps(data))
    print(response.json()["choices"][0]["message"]["content"])

Final memory usage (6538 MB):

vLLM serve command (using V0):

VLLM_USE_V1=0 vllm serve mistralai/Pixtral-12B-2409 --tokenizer_mode mistral --limit_mm_per_prompt 'image=4' --disable-mm-preprocessor-cache

Initial memory usage (8845 MB):

Final memory usage -after 100 image inference requests- (8835 MB):

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.

[ywang96]

Hey @oyerli - thanks for reporting this! As a quick test to help us verifify where the issue is, can you try specifying --disable-mm-preprocessor-cache and see if the issue still persists?

[WoosukKwon]

@ywang96 I think he already used it?

vllm serve mistralai/Pixtral-12B-2409 --tokenizer_mode mistral --limit_mm_per_prompt 'image=4' --disable-mm-preprocessor-cache

[oyerli]

Hey @oyerli - thanks for reporting this! As a quick test to help us verifify where the issue is, can you try specifying --disable-mm-preprocessor-cache and see if the issue still persists?

Thanks for the quick follow-up! We already had --disable-mm-preprocessor-cache enabled in our runs — you can see it included in the original serve commands in the bug report. Unfortunately, the memory leak still occurs with that flag set. Let me know if there's anything else you'd like us to try — happy to help!

[ywang96]

Sorry I missed that in your command, so this is something different from the previous fix we added for 0.8.1. I’ll look into this today!

[ywang96]

Hey @oyerli - I just got a chance to test the script but I couldn't repro this from our main branch - can you give the main branch a try and see if the issue persists for you?

In the mean time, I'm also going to try our official 0.8.1 PyPI package and see if I can repro.

EDIT: Confirmed this is reprocible on 0.8.1 - I will try to bisect and find the culprit, but please let me know if main branch has fixed this for you production traffic!

[oyerli]

Hey @oyerli - I just got a chance to test the script but I couldn't repro this from our main branch - can you give the main branch a try and see if the issue persists for you?

In the mean time, I'm also going to try our official 0.8.1 PyPI package and see if I can repro.

EDIT: Confirmed this is reprocible on 0.8.1 - I will try to bisect and find the culprit, but please let me know if main branch has fixed this for you production traffic!

Thank you @ywang96. I just confirmed that the main branch does indeed resolve the memory leak on our end as well.

I did a bisect and found that commit d8e82bc is what fixes the issue.

It's a relatively simple change, but I’d appreciate any insight into why this was causing the leak—especially if it might be tied to our usage pattern, so we can avoid similar issues going forward.

[robertgshaw2-redhat]

Hey @oyerli - I just got a chance to test the script but I couldn't repro this from our main branch - can you give the main branch a try and see if the issue persists for you?
In the mean time, I'm also going to try our official 0.8.1 PyPI package and see if I can repro.
EDIT: Confirmed this is reprocible on 0.8.1 - I will try to bisect and find the culprit, but please let me know if main branch has fixed this for you production traffic!

Thank you @ywang96. I just confirmed that the main branch does indeed resolve the memory leak on our end as well.

I did a bisect and found that commit d8e82bc is what fixes the issue.

It's a relatively simple change, but I’d appreciate any insight into why this was causing the leak—especially if it might be tied to our usage pattern, so we can avoid similar issues going forward.

@oyerli - there is nothing that you were doing wrong, just a last minute change we made to 0.8.0/1 which introduced a bug.

We have the following code:

if not self.scheduler.has_requests():
    return EngineCoreOutputs(
        outputs=[],
        scheduler_stats=self.scheduler.make_stats(),
    )
scheduler_output = self.scheduler.schedule()

# This case may occur when the only unfinished requests are
# structured output requests where the grammar has not finished
# compiling yet, so there's nothing to run.
if scheduler_output.total_num_scheduled_tokens == 0:
    return EngineCoreOutputs(
        outputs=[],
        scheduler_stats=self.scheduler.make_stats(),
    )

output = self.model_executor.execute_model(scheduler_output)
engine_core_outputs = self.scheduler.update_from_output(
    scheduler_output, output)  # type: ignore

return engine_core_outputs

What happens is:

• The previous iteration of scheduler.update_from_output puts requests into scheduler.finished. This frees the req from the POV of the scheduler
• self.scheduler.has_requests() is True if there are requests in scheduler.finished
• scheduler.schedule() resets scheduler.finished = set(), adds finished to scheduler_output.finished
• self.model_executor.execute_model(scheduler_output) frees the requests from `scheduler_output.finished in the persistent batch in the model runner. So, execute model is responsible for freeing the requests from the POV of the persistent batch << this is where the memory leak is happening

This means that if a request finishes and no other requests are running, we were leaking in the persistent batch.

We added the optimization to skip execute_model when scheduler_output.total_num_scheduled_tokens == 0 as part of the guided decoding work which landed right before the release. This was just a miss on our part. Its exacerbated by image data since there is more data per request.

Thank you for reporting the issue. We really apologize for the issue it may have caused you. We will be doing a new release on Monday.

Thanks! Rob