[Bug]: Memory profiler does not consider CUDA context memory #12059
Description
Your current environment
The output of `python collect_env.py`
Collecting environment information...
PyTorch version: 2.5.1+cu124
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A
OS: Ubuntu 24.04.1 LTS (x86_64)
GCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Clang version: Could not collect
CMake version: version 3.28.3
Libc version: glibc-2.39
Python version: 3.12.3 (main, Nov 6 2024, 18:32:19) [GCC 13.2.0] (64-bit runtime)
Python platform: Linux-6.8.0-51-generic-x86_64-with-glibc2.39
Is CUDA available: True
CUDA runtime version: 12.6.68
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 4090
Nvidia driver version: 560.35.03
cuDNN version: Could not collect
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, 48 bits virtual
Byte Order: Little Endian
CPU(s): 32
On-line CPU(s) list: 0-31
Vendor ID: GenuineIntel
Model name: 13th Gen Intel(R) Core(TM) i9-13900K
CPU family: 6
Model: 183
Thread(s) per core: 2
Core(s) per socket: 24
Socket(s): 1
Stepping: 1
CPU(s) scaling MHz: 66%
CPU max MHz: 5800.0000
CPU min MHz: 800.0000
BogoMIPS: 5990.40
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 sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdseed adx smap clflushopt clwb intel_pt sha_ni xsaveopt xsavec xgetbv1 xsaves split_lock_detect user_shstk avx_vnni dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp hwp_pkg_req hfi vnmi umip pku ospke waitpkg gfni vaes vpclmulqdq tme rdpid movdiri movdir64b fsrm md_clear serialize pconfig arch_lbr ibt flush_l1d arch_capabilities
Virtualization: VT-x
L1d cache: 896 KiB (24 instances)
L1i cache: 1.3 MiB (24 instances)
L2 cache: 32 MiB (12 instances)
L3 cache: 36 MiB (1 instance)
NUMA node(s): 1
NUMA node0 CPU(s): 0-31
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: Mitigation; Clear Register File
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; 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-ml-py==12.560.30
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] pyzmq==26.2.0
[pip3] torch==2.5.1
[pip3] torchvision==0.20.1
[pip3] transformers==4.48.0
[pip3] triton==3.1.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.6.6.post2.dev211+g2e0e0176
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 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
LD_LIBRARY_PATH=/home/benchislett/Repos/vllm/.venv/lib/python3.12/site-packages/cv2/../../lib64::/usr/local/cuda-12/lib64:/usr/local/cuda-12/targets/x86_64-linux/lib:/usr/lib:/usr/local/cuda-12/lib64:/usr/local/cuda-12/targets/x86_64-linux/lib:/usr/lib:/usr/local/cuda-12/lib64:/usr/local/cuda-12/targets/x86_64-linux/lib:/usr/lib:/usr/local/cuda-12/lib64:/usr/local/cuda-12/targets/x86_64-linux/lib:/usr/lib
CUDA_MODULE_LOADING=LAZY
Model Input Dumps
No response
🐛 Describe the bug
Issue #10451 highlights a problem in previous memory profiling strategy, and #10511 overhauls the profiler to return the old behavior.
Current vLLM takes an initial measurement in Worker.init_device() that calculates self.init_gpu_memory (worker.py#L143), and uses this as a baseline to calculate the amount of non-torch memory used.
The function of this baseline is to be used in the memory profiling as described in utils.py#L1971:
baseline_memory_in_bytes: memory used by all the components other than
the current vLLM instance. It contains: memory used by other processes, memory
used by another vLLM instance in the same process, etc. It is usually measured
before the current vLLM instance initialize the device. And we assume it is
constant during the profiling of the current vLLM instance.
Using torch.cuda.mem_get_info() to compute this baseline violates this assumption. This is because it includes the non-torch CUDA context memory that is allocated when the cuda runtime API is used for the first time in each process.
This is very easy to check. First, check how much memory is currently used with nvidia-smi, and then run the following:
import torch
free, total = torch.cuda.mem_get_info()
print(f"Total GPU memory usage: {(total - free) / (1024**2):.2f} MB")The exact discrepancy depends on the GPU type and is hard to measure. I found mention of this quantity in an old CUDA forums post. This overhead seems very difficult to measure automatically at run-time, as any CUDA runtime API will already have allocated this overhead, and NVML can only introspect this usage using PID information which is not reliable in all cases (such as in containerised environments.
On my 4090, it measures 384 MiB. On an H100, it's around 520 MiB.
The best solution I can think of is to account for this in the baseline:
self.init_gpu_memory = max(0, self.init_gpu_memory - (0.5 * GiB))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.