[Bug]: RAM OOM Error Loading 480GB MoE Model Despite Fix in PR #1395

[hxer7963]

Your current environment

PyTorch version: 2.1.2+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.3 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: 14.0.0-1ubuntu1.1
CMake version: version 3.26.2
Libc version: glibc-2.35

Python version: 3.9.18 (main, Sep 11 2023, 13:41:44) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.4.119-19-0009.11-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.3.107
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA A800-SXM4-80GB
GPU 1: NVIDIA A800-SXM4-80GB
GPU 2: NVIDIA A800-SXM4-80GB
GPU 3: NVIDIA A800-SXM4-80GB
GPU 4: NVIDIA A800-SXM4-80GB
GPU 5: NVIDIA A800-SXM4-80GB
GPU 6: NVIDIA A800-SXM4-80GB
GPU 7: NVIDIA A800-SXM4-80GB

Nvidia driver version: 470.182.03
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.0.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.0.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.0.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.0.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.0.0
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.0.0
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.0.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.0.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.7
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.7
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: 48 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 232
On-line CPU(s) list: 0-231
Vendor ID: AuthenticAMD
Model name: AMD EPYC 7K83 64-Core Processor
CPU family: 25
Model: 1
Thread(s) per core: 2
Core(s) per socket: 58
Socket(s): 2
Stepping: 1
BogoMIPS: 4890.80
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 amd_dcm tsc_known_freq pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw topoext perfctr_core invpcid_single ssbd ibrs ibpb stibp vmmcall fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves clzero xsaveerptr wbnoinvd arat umip pku ospke vaes vpclmulqdq rdpid fsrm
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 3.6 MiB (116 instances)
L1i cache: 3.6 MiB (116 instances)
L2 cache: 58 MiB (116 instances)
L3 cache: 512 MiB (16 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-115
NUMA node1 CPU(s): 116-231
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: 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; Full AMD retpoline, IBPB conditional, IBRS_FW, STIBP conditional, RSB filling
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected

Versions of relevant libraries:
[pip3] mypy==0.991
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.4
[pip3] nvidia-nccl-cu12==2.18.1
[pip3] torch==2.1.2
[pip3] triton==2.1.0
[conda] numpy 1.26.4 pypi_0 pypi
[conda] nvidia-nccl-cu12 2.18.1 pypi_0 pypi
[conda] torch 2.1.2 pypi_0 pypi
[conda] triton 2.1.0 pypi_0 pypiROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.4.0
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 CPU Affinity NUMA Affinity
GPU0 X NV8 NV8 NV8 NV8 NV8 NV8 NV8 0-115 0
GPU1 NV8 X NV8 NV8 NV8 NV8 NV8 NV8 0-115 0
GPU2 NV8 NV8 X NV8 NV8 NV8 NV8 NV8 0-115 0
GPU3 NV8 NV8 NV8 X NV8 NV8 NV8 NV8 0-115 0
GPU4 NV8 NV8 NV8 NV8 X NV8 NV8 NV8 116-231 1
GPU5 NV8 NV8 NV8 NV8 NV8 X NV8 NV8 116-231 1
GPU6 NV8 NV8 NV8 NV8 NV8 NV8 X NV8 116-231 1
GPU7 NV8 NV8 NV8 NV8 NV8 NV8 NV8 X 116-231 1

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

🐛 Describe the bug

I run the following command on a server with 8 * A800 (80GB) GPUs and 1.8TB of memory:

python -m vllm.entrypoints.openai.api_server --model /mnt/llm_dataset/models/xverse_model/moe/moe24b-v112-276b/  --dtype auto --api-key xxx
 --trust-remote-code -tp 8 --port 8001

I noticed that vLLM initially loads the model into the 8 A800 GPUs, with each GPU consuming approximately 64GB of memory on average, without pre-allocating memory. Subsequently, I observed that the DRAM memory keeps increasing until an Out-of-Memory (OOM) error occurs. Please refer to the screenshot below for details.

Then, I learned that I can reduce the risk of RAM OOM by setting the --max-parallel-loading-workers parameter. However, when I set --max-parallel-loading-workers 1, I encountered the following error:

NotImplementedError: max_concurrent_workers is not supported yet.

#2588

I noticed that PR #1395 has fixed the RAM OOM issue, but why am I still encountering OOM errors when loading a 480GB MoE model with 8tp and 1.8TB of memory?

[mgoin]

Hi @hxer7963, if my math is correct I do not think you can fit an unquantized 480GB model on 8x80GB GPUs. 8 * 80 = 640GB available memory and the model weights at BF16 would take up 2 * 480 = 960GB, so that is very much over the memory available. Maybe if you had an 8bit or 4bit quantization with GPTQ, you could fit the model.

Also, I would recommend trying to upgrade vLLM as you are on 0.4.0 per vLLM Version: 0.4.0