Issue Station
Sign In
Tools hereToolbox

vllm - ✅(Solved) Fix [Bug]: Prefetch CPU offload OOMs; `VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY` must be implemented [2 pull requests, 2 comments, 3 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
vllm-project/vllm#37672•Fetched 2026-04-08 01:04:10
View on GitHub
Comments
2
Participants
3
Timeline
11
Reactions
0
Author
ehfdehfd
Participants
ehfdehfdhe-yufenghe-yufengwzhao18wzhao18
Timeline (top)
commented ×2cross-referenced ×2mentioned ×2subscribed ×2

Fix Action

Fix / Workaround

============================== CPU Info

Architecture: aarch64 CPU op-mode(s): 64-bit Byte Order: Little Endian CPU(s): 72 On-line CPU(s) list: 0-71 Vendor ID: ARM Model name: Neoverse-V2 Model: 0 Thread(s) per core: 1 Core(s) per cluster: 72 Socket(s): - Cluster(s): 1 Stepping: r0p0 Frequency boost: disabled CPU max MHz: 3447.0000 CPU min MHz: 81.0000 BogoMIPS: 2000.00 Flags: fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 sm3 sm4 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm ssbs sb paca pacg dcpodp sve2 sveaes svepmull svebitperm svesha3 svesm4 flagm2 frint svei8mm svebf16 i8mm bf16 dgh bti L1d cache: 4.5 MiB (72 instances) L1i cache: 4.5 MiB (72 instances) L2 cache: 72 MiB (72 instances) L3 cache: 114 MiB (1 instance) NUMA node(s): 9 NUMA node0 CPU(s): 0-71 NUMA node1 CPU(s): NUMA node2 CPU(s): NUMA node3 CPU(s): NUMA node4 CPU(s): NUMA node5 CPU(s): NUMA node6 CPU(s): NUMA node7 CPU(s): NUMA node8 CPU(s): 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 Vulnerability Spectre v1: Mitigation; __user pointer sanitization Vulnerability Spectre v2: Not affected Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected

PR fix notes

PR #32993: [Feature] Support CPU Offloading without Pytorch Pinned Memory that leads to doubled allocation

Description (problem / solution / changelog)

Purpose

Background: Current PyTorch pin_memory implementation rounds allocation to the next power of 2 (see Issue #150517, PR #171662). In the worst case, this results in 2x the expected CPU memory usage. For instance, running nvidia/Kimi-K2-Thinking-NVFP4 on a system with one GB300 GPU and 500 GB of CPU memory results in OOM, even though sufficient memory is available.

Script to demonstrate the issue:

import torch

import sys
import psutil
gb_size = int(sys.argv[1])

print(f"Allocating {gb_size} GB")
before = psutil.virtual_memory().used
torch.empty(gb_size * 1024 ** 3, dtype=torch.uint8, device="cpu", pin_memory=True)

after = psutil.virtual_memory().used
print(f"Memory used: {(after - before) / 1024 ** 3} GB")

Allocating 257 GB of pinned memory leads to 512 GB CPU allocation.

Changes: This PR provides a workaround to get around using Pytorch pin memory to use UVA-based (unified virtual addressing) CPU offloading. In addition, it fixes various issues on the non-UVA offloading code path (explicit transfer between CPU and GPU).

There is a pytorch PR that is looking to address the doubled pin memory usage issue. This PR provides a temporary workaround for the problem. The workaround is enabled through two environment variables:

  • VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY: disable using pytorch pin memory in CPU offloading
  • VLLM_WEIGHT_OFFLOADING_DISABLE_UVA: disable using UVA in CPU offloading

When neither env vars are set, the original offloading logic (pytorch pinned memory + UVA) is used.

Test Plan

Added testing for the env-var enabled code paths in tests/basic_correctness/test_cpu_offload.py.

Also tested nvidia/Kimi-K2-Thinking-NVFP4 on one GB300 GPU and 500 GB CPU memory:

On current main:

python3 -m vllm.entrypoints.openai.api_server \
	--model nvidia/Kimi-K2-Thinking-NVFP4 \
	--trust-remote-code \
        --cpu-offload-gb 350
File "/vllm/model_executor/models/utils.py", line 707, in make_layers
^[[core.py:935]     maybe_offload_to_cpu(layer_fn(prefix=f"{prefix}.{idx}"))
 [core.py:935]   File "/vllm/model_executor/models/utils.py", line 652, in maybe_offload_to_cpu
 [core.py:935]     cpu_data = torch.empty_strided(
 [core.py:935]                ^^^^^^^^^^^^^^^^^^^^
 [core.py:935]   File "/vllm/.venv/lib/python3.12/site-packages/torch/utils/_device.py", line 109, in __torch_function__
 [core.py:935]     return func(*args, **kwargs)
 [core.py:935]            ^^^^^^^^^^^^^^^^^^^^^
 [core.py:935] torch.AcceleratorError: CUDA error: out of memory

Test Result

With the changes, the following command will work.

VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY=1 \
python3 -m vllm.entrypoints.openai.api_server \
	--model nvidia/Kimi-K2-Thinking-NVFP4 \
	--trust-remote-code \
        --cpu-offload-gb 350
[gpu_model_runner.py:3947] Model loading took 199.97 GiB memory and 173.309022 seconds
<details> <summary> Essential Elements of an Effective PR Description Checklist </summary>
  • The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
  • The test plan, such as providing test command.
  • The test results, such as pasting the results comparison before and after, or e2e results
  • (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
  • (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.
</details>

Changed files

Code Example

Collecting environment information...
==============================
        System Info
==============================
OS                           : Ubuntu 22.04.5 LTS (aarch64)
GCC version                  : (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0
Clang version                : Could not collect
CMake version                : Could not collect
Libc version                 : glibc-2.35

==============================
       PyTorch Info
==============================
PyTorch version              : 2.10.0+cu129
Is debug build               : False
CUDA used to build PyTorch   : 12.9
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.12.13 (main, Mar  4 2026, 09:23:07) [GCC 11.4.0] (64-bit runtime)
Python platform              : Linux-6.8.0-1041-nvidia-64k-aarch64-with-glibc2.35

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : 12.9.86
CUDA_MODULE_LOADING set to   :
GPU models and configuration : GPU 0: NVIDIA GH200 480GB
Nvidia driver version        : 580.95.05
cuDNN version                : Could not collect
HIP runtime version          : N/A
MIOpen runtime version       : N/A
Is XNNPACK available         : True

==============================
          CPU Info
==============================
Architecture:                         aarch64
CPU op-mode(s):                       64-bit
Byte Order:                           Little Endian
CPU(s):                               72
On-line CPU(s) list:                  0-71
Vendor ID:                            ARM
Model name:                           Neoverse-V2
Model:                                0
Thread(s) per core:                   1
Core(s) per cluster:                  72
Socket(s):                            -
Cluster(s):                           1
Stepping:                             r0p0
Frequency boost:                      disabled
CPU max MHz:                          3447.0000
CPU min MHz:                          81.0000
BogoMIPS:                             2000.00
Flags:                                fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 sm3 sm4 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm ssbs sb paca pacg dcpodp sve2 sveaes svepmull svebitperm svesha3 svesm4 flagm2 frint svei8mm svebf16 i8mm bf16 dgh bti
L1d cache:                            4.5 MiB (72 instances)
L1i cache:                            4.5 MiB (72 instances)
L2 cache:                             72 MiB (72 instances)
L3 cache:                             114 MiB (1 instance)
NUMA node(s):                         9
NUMA node0 CPU(s):                    0-71
NUMA node1 CPU(s):
NUMA node2 CPU(s):
NUMA node3 CPU(s):
NUMA node4 CPU(s):
NUMA node5 CPU(s):
NUMA node6 CPU(s):
NUMA node7 CPU(s):
NUMA node8 CPU(s):
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
Vulnerability Spectre v1:             Mitigation; __user pointer sanitization
Vulnerability Spectre v2:             Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

==============================
Versions of relevant libraries
==============================
[pip3] flashinfer-python==0.6.6
[pip3] numpy==2.2.6
[pip3] nvidia-cublas-cu12==12.9.1.4
[pip3] nvidia-cuda-cupti-cu12==12.9.79
[pip3] nvidia-cuda-nvrtc-cu12==12.9.86
[pip3] nvidia-cuda-runtime-cu12==12.9.79
[pip3] nvidia-cudnn-cu12==9.10.2.21
[pip3] nvidia-cudnn-frontend==1.18.0
[pip3] nvidia-cufft-cu12==11.4.1.4
[pip3] nvidia-cufile-cu12==1.14.1.1
[pip3] nvidia-curand-cu12==10.3.10.19
[pip3] nvidia-cusolver-cu12==11.7.5.82
[pip3] nvidia-cusparse-cu12==12.5.10.65
[pip3] nvidia-cusparselt-cu12==0.7.1
[pip3] nvidia-cutlass-dsl==4.4.2
[pip3] nvidia-cutlass-dsl-libs-base==4.4.2
[pip3] nvidia-ml-py==13.590.48
[pip3] nvidia-nccl-cu12==2.27.5
[pip3] nvidia-nvjitlink-cu12==12.9.86
[pip3] nvidia-nvshmem-cu12==3.4.5
[pip3] nvidia-nvtx-cu12==12.9.79
[pip3] pyzmq==27.1.0
[pip3] torch==2.10.0+cu129
[pip3] torch_c_dlpack_ext==0.1.5
[pip3] torchaudio==2.10.0+cu129
[pip3] torchvision==0.25.0+cu129
[pip3] transformers==4.57.6
[pip3] triton==3.6.0
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.17.2rc1.dev96+ge3126cd10 (git sha: e3126cd10)
vLLM Build Flags:
  CUDA Archs: 8.7 8.9 9.0 10.0+PTX 12.0; ROCm: Disabled
GPU Topology:
        GPU0    NIC0    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      NODE    0-71    0               1
NIC0    NODE     X

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

NIC Legend:

  NIC0: rocep1s0f1

==============================
     Environment Variables
==============================
NVIDIA_VISIBLE_DEVICES=GPU-9302ac7f-9fb3-24c3-e0d1-7c60f4cf076d
NVIDIA_REQUIRE_CUDA=cuda>=12.9 brand=unknown,driver>=535,driver<536 brand=grid,driver>=535,driver<536 brand=tesla,driver>=535,driver<536 brand=nvidia,driver>=535,driver<536 brand=quadro,driver>=535,driver<536 brand=quadrortx,driver>=535,driver<536 brand=nvidiartx,driver>=535,driver<536 brand=vapps,driver>=535,driver<536 brand=vpc,driver>=535,driver<536 brand=vcs,driver>=535,driver<536 brand=vws,driver>=535,driver<536 brand=cloudgaming,driver>=535,driver<536 brand=unknown,driver>=550,driver<551 brand=grid,driver>=550,driver<551 brand=tesla,driver>=550,driver<551 brand=nvidia,driver>=550,driver<551 brand=quadro,driver>=550,driver<551 brand=quadrortx,driver>=550,driver<551 brand=nvidiartx,driver>=550,driver<551 brand=vapps,driver>=550,driver<551 brand=vpc,driver>=550,driver<551 brand=vcs,driver>=550,driver<551 brand=vws,driver>=550,driver<551 brand=cloudgaming,driver>=550,driver<551 brand=unknown,driver>=560,driver<561 brand=grid,driver>=560,driver<561 brand=tesla,driver>=560,driver<561 brand=nvidia,driver>=560,driver<561 brand=quadro,driver>=560,driver<561 brand=quadrortx,driver>=560,driver<561 brand=nvidiartx,driver>=560,driver<561 brand=vapps,driver>=560,driver<561 brand=vpc,driver>=560,driver<561 brand=vcs,driver>=560,driver<561 brand=vws,driver>=560,driver<561 brand=cloudgaming,driver>=560,driver<561 brand=unknown,driver>=565,driver<566 brand=grid,driver>=565,driver<566 brand=tesla,driver>=565,driver<566 brand=nvidia,driver>=565,driver<566 brand=quadro,driver>=565,driver<566 brand=quadrortx,driver>=565,driver<566 brand=nvidiartx,driver>=565,driver<566 brand=vapps,driver>=565,driver<566 brand=vpc,driver>=565,driver<566 brand=vcs,driver>=565,driver<566 brand=vws,driver>=565,driver<566 brand=cloudgaming,driver>=565,driver<566 brand=unknown,driver>=570,driver<571 brand=grid,driver>=570,driver<571 brand=tesla,driver>=570,driver<571 brand=nvidia,driver>=570,driver<571 brand=quadro,driver>=570,driver<571 brand=quadrortx,driver>=570,driver<571 brand=nvidiartx,driver>=570,driver<571 brand=vapps,driver>=570,driver<571 brand=vpc,driver>=570,driver<571 brand=vcs,driver>=570,driver<571 brand=vws,driver>=570,driver<571 brand=cloudgaming,driver>=570,driver<571
TORCH_CUDA_ARCH_LIST=8.7 8.9 9.0 10.0+PTX 12.0
PYTORCH_ALLOC_CONF=expandable_segments:True
NVIDIA_DRIVER_CAPABILITIES=compute,utility
VLLM_USAGE_SOURCE=production-docker-image
CUDA_VERSION=12.9.1
VLLM_WEIGHT_OFFLOADING_DISABLE_UVA=1
VLLM_ENABLE_CUDA_COMPATIBILITY=0
LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:/usr/local/cuda/lib64
VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY=1
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_root
RAW_BUFFERClick to expand / collapse

Your current environment

<details> <summary>The output of <code>python collect_env.py</code></summary> 
Collecting environment information...
==============================
        System Info
==============================
OS                           : Ubuntu 22.04.5 LTS (aarch64)
GCC version                  : (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0
Clang version                : Could not collect
CMake version                : Could not collect
Libc version                 : glibc-2.35

==============================
       PyTorch Info
==============================
PyTorch version              : 2.10.0+cu129
Is debug build               : False
CUDA used to build PyTorch   : 12.9
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.12.13 (main, Mar  4 2026, 09:23:07) [GCC 11.4.0] (64-bit runtime)
Python platform              : Linux-6.8.0-1041-nvidia-64k-aarch64-with-glibc2.35

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : 12.9.86
CUDA_MODULE_LOADING set to   :
GPU models and configuration : GPU 0: NVIDIA GH200 480GB
Nvidia driver version        : 580.95.05
cuDNN version                : Could not collect
HIP runtime version          : N/A
MIOpen runtime version       : N/A
Is XNNPACK available         : True

==============================
          CPU Info
==============================
Architecture:                         aarch64
CPU op-mode(s):                       64-bit
Byte Order:                           Little Endian
CPU(s):                               72
On-line CPU(s) list:                  0-71
Vendor ID:                            ARM
Model name:                           Neoverse-V2
Model:                                0
Thread(s) per core:                   1
Core(s) per cluster:                  72
Socket(s):                            -
Cluster(s):                           1
Stepping:                             r0p0
Frequency boost:                      disabled
CPU max MHz:                          3447.0000
CPU min MHz:                          81.0000
BogoMIPS:                             2000.00
Flags:                                fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 sm3 sm4 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm ssbs sb paca pacg dcpodp sve2 sveaes svepmull svebitperm svesha3 svesm4 flagm2 frint svei8mm svebf16 i8mm bf16 dgh bti
L1d cache:                            4.5 MiB (72 instances)
L1i cache:                            4.5 MiB (72 instances)
L2 cache:                             72 MiB (72 instances)
L3 cache:                             114 MiB (1 instance)
NUMA node(s):                         9
NUMA node0 CPU(s):                    0-71
NUMA node1 CPU(s):
NUMA node2 CPU(s):
NUMA node3 CPU(s):
NUMA node4 CPU(s):
NUMA node5 CPU(s):
NUMA node6 CPU(s):
NUMA node7 CPU(s):
NUMA node8 CPU(s):
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
Vulnerability Spectre v1:             Mitigation; __user pointer sanitization
Vulnerability Spectre v2:             Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

==============================
Versions of relevant libraries
==============================
[pip3] flashinfer-python==0.6.6
[pip3] numpy==2.2.6
[pip3] nvidia-cublas-cu12==12.9.1.4
[pip3] nvidia-cuda-cupti-cu12==12.9.79
[pip3] nvidia-cuda-nvrtc-cu12==12.9.86
[pip3] nvidia-cuda-runtime-cu12==12.9.79
[pip3] nvidia-cudnn-cu12==9.10.2.21
[pip3] nvidia-cudnn-frontend==1.18.0
[pip3] nvidia-cufft-cu12==11.4.1.4
[pip3] nvidia-cufile-cu12==1.14.1.1
[pip3] nvidia-curand-cu12==10.3.10.19
[pip3] nvidia-cusolver-cu12==11.7.5.82
[pip3] nvidia-cusparse-cu12==12.5.10.65
[pip3] nvidia-cusparselt-cu12==0.7.1
[pip3] nvidia-cutlass-dsl==4.4.2
[pip3] nvidia-cutlass-dsl-libs-base==4.4.2
[pip3] nvidia-ml-py==13.590.48
[pip3] nvidia-nccl-cu12==2.27.5
[pip3] nvidia-nvjitlink-cu12==12.9.86
[pip3] nvidia-nvshmem-cu12==3.4.5
[pip3] nvidia-nvtx-cu12==12.9.79
[pip3] pyzmq==27.1.0
[pip3] torch==2.10.0+cu129
[pip3] torch_c_dlpack_ext==0.1.5
[pip3] torchaudio==2.10.0+cu129
[pip3] torchvision==0.25.0+cu129
[pip3] transformers==4.57.6
[pip3] triton==3.6.0
[conda] Could not collect

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
vLLM Version                 : 0.17.2rc1.dev96+ge3126cd10 (git sha: e3126cd10)
vLLM Build Flags:
  CUDA Archs: 8.7 8.9 9.0 10.0+PTX 12.0; ROCm: Disabled
GPU Topology:
        GPU0    NIC0    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      NODE    0-71    0               1
NIC0    NODE     X

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

NIC Legend:

  NIC0: rocep1s0f1

==============================
     Environment Variables
==============================
NVIDIA_VISIBLE_DEVICES=GPU-9302ac7f-9fb3-24c3-e0d1-7c60f4cf076d
NVIDIA_REQUIRE_CUDA=cuda>=12.9 brand=unknown,driver>=535,driver<536 brand=grid,driver>=535,driver<536 brand=tesla,driver>=535,driver<536 brand=nvidia,driver>=535,driver<536 brand=quadro,driver>=535,driver<536 brand=quadrortx,driver>=535,driver<536 brand=nvidiartx,driver>=535,driver<536 brand=vapps,driver>=535,driver<536 brand=vpc,driver>=535,driver<536 brand=vcs,driver>=535,driver<536 brand=vws,driver>=535,driver<536 brand=cloudgaming,driver>=535,driver<536 brand=unknown,driver>=550,driver<551 brand=grid,driver>=550,driver<551 brand=tesla,driver>=550,driver<551 brand=nvidia,driver>=550,driver<551 brand=quadro,driver>=550,driver<551 brand=quadrortx,driver>=550,driver<551 brand=nvidiartx,driver>=550,driver<551 brand=vapps,driver>=550,driver<551 brand=vpc,driver>=550,driver<551 brand=vcs,driver>=550,driver<551 brand=vws,driver>=550,driver<551 brand=cloudgaming,driver>=550,driver<551 brand=unknown,driver>=560,driver<561 brand=grid,driver>=560,driver<561 brand=tesla,driver>=560,driver<561 brand=nvidia,driver>=560,driver<561 brand=quadro,driver>=560,driver<561 brand=quadrortx,driver>=560,driver<561 brand=nvidiartx,driver>=560,driver<561 brand=vapps,driver>=560,driver<561 brand=vpc,driver>=560,driver<561 brand=vcs,driver>=560,driver<561 brand=vws,driver>=560,driver<561 brand=cloudgaming,driver>=560,driver<561 brand=unknown,driver>=565,driver<566 brand=grid,driver>=565,driver<566 brand=tesla,driver>=565,driver<566 brand=nvidia,driver>=565,driver<566 brand=quadro,driver>=565,driver<566 brand=quadrortx,driver>=565,driver<566 brand=nvidiartx,driver>=565,driver<566 brand=vapps,driver>=565,driver<566 brand=vpc,driver>=565,driver<566 brand=vcs,driver>=565,driver<566 brand=vws,driver>=565,driver<566 brand=cloudgaming,driver>=565,driver<566 brand=unknown,driver>=570,driver<571 brand=grid,driver>=570,driver<571 brand=tesla,driver>=570,driver<571 brand=nvidia,driver>=570,driver<571 brand=quadro,driver>=570,driver<571 brand=quadrortx,driver>=570,driver<571 brand=nvidiartx,driver>=570,driver<571 brand=vapps,driver>=570,driver<571 brand=vpc,driver>=570,driver<571 brand=vcs,driver>=570,driver<571 brand=vws,driver>=570,driver<571 brand=cloudgaming,driver>=570,driver<571
TORCH_CUDA_ARCH_LIST=8.7 8.9 9.0 10.0+PTX 12.0
PYTORCH_ALLOC_CONF=expandable_segments:True
NVIDIA_DRIVER_CAPABILITIES=compute,utility
VLLM_USAGE_SOURCE=production-docker-image
CUDA_VERSION=12.9.1
VLLM_WEIGHT_OFFLOADING_DISABLE_UVA=1
VLLM_ENABLE_CUDA_COMPATIBILITY=0
LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:/usr/local/cuda/lib64
VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY=1
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
TORCHINDUCTOR_CACHE_DIR=/tmp/torchinductor_root
</details>

🐛 Describe the bug

Environment: GH200

Command: python3 -m vllm.entrypoints.openai.api_server --port 5000 --host 0.0.0.0 --download-dir /workspace/.cache/huggingface/hub --model zai-org/GLM-4.7-FP8 --api-server-count 8 --tensor-parallel-size 1 --trust-remote-code --enable-chunked-prefill --enable-prefix-caching --max-num-seqs 32 --gpu-memory-utilization 0.75 --max-model-len 202752 --enable-auto-tool-choice --tool-call-parser glm47 --reasoning-parser glm45 --offload-backend prefetch --offload-group-size 1 --offload-num-in-group 1 --offload-prefetch-step 1

<img width="1366" height="494" alt="Image" src="https://github.com/user-attachments/assets/af51ae0b-62c6-47bc-9e9d-58576dc6e625" />

OOM (at 474 GiB, maximum of a GH200) occurs in GLM-4.7-FP8 (https://huggingface.co/zai-org/GLM-4.7-FP8, Model weight = 362 GB = 338 GiB) in the same way described in https://github.com/vllm-project/vllm/pull/32993. And looking at the code, there are no references of VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY present in vllm/model_executor/offloader/prefetch.py or vllm/model_executor/offloader/prefetch_ops.py

CC @wzhao18

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.

extent analysis

Fix Plan

To address the OOM issue with the GLM-4.7-FP8 model on a GH200 GPU, we need to optimize memory utilization. Given the model's size (338 GiB) and the GPU's memory (480 GB), but considering the OOM occurs at 474 GiB, we'll focus on optimizing the VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY and VLLM_WEIGHT_OFFLOADING_DISABLE_UVA environment variables, as well as model loading and offloading strategies.

  1. Environment Variable Adjustments:

    • Set VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY=0 to enable pinning memory, which can help reduce memory fragmentation and utilization.
    • Ensure VLLM_WEIGHT_OFFLOADING_DISABLE_UVA=1 to disable UVA (Unified Virtual Addressing) for potential memory savings, though this might impact performance.

  2. Model Loading and Offloading:

    • Review and adjust the offload-backend, offload-group-size, and offload-num-in-group parameters to better manage how the model is loaded and offloaded from the GPU memory.
    • Consider reducing --max-num-seqs or --gpu-memory-utilization to lower the memory demand.

  3. Code Adjustments:

    • In vllm/model_executor/offloader/prefetch.py and vllm/model_executor/offloader/prefetch_ops.py, add logic to respect and utilize the VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY environment variable for more efficient memory management.

Example adjustments to the command:

VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY=0 VLLM_WEIGHT_OFFLOADING_DISABLE_UVA=1 python3 -m vllm.entrypoints.openai.api_server \
--port 5000 --host 0.0.0.0 --download-dir /workspace/.cache/huggingface/hub \
--model zai-org/GLM-4.7-FP8 --api-server-count 8 --tensor-parallel-size 1 \
--trust-remote-code --enable-chunked-prefill --enable-prefix-caching --max-num-seqs 16 \
--gpu-memory-utilization 0.6 --max-model-len 202752 --enable-auto-tool-choice \
--tool-call-parser glm47 --reasoning-parser glm45 --offload-backend prefetch \
--offload-group-size 2 --offload-num-in-group 2 --offload-prefetch-step 2

Verification

  • Monitor GPU memory usage during model execution to ensure it stays below the OOM threshold.
  • Test with different max-num-seqs and gpu-memory-utilization settings to find an optimal balance between performance and memory usage.

Extra Tips

  • Regularly update NVIDIA drivers and CUDA versions to ensure compatibility and performance.
  • Consider using

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…
#api#ssr#installation#tensor shape#device allocation#model download#tokenizer error#environment variable

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING

vllm - ✅(Solved) Fix [Bug]: Prefetch CPU offload OOMs; `VLLM_WEIGHT_OFFLOADING_DISABLE_PIN_MEMORY` must be implemented [2 pull requests, 2 comments, 3 participants]