[Bug]: vllm:request_inference_time_seconds_bucket has too few buckets for long inference requests

[smarterclayton]

Your current environment

The output of `python collect_env.py`

INFO 03-19 22:07:01 [__init__.py:256] Automatically detected platform cuda.
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 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.31.6
Libc version: glibc-2.35

Python version: 3.12.9 (main, Mar 17 2025, 21:01:58) [Clang 20.1.0 ] (64-bit runtime)
Python platform: Linux-6.6.72+-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.4.131
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA H100 80GB HBM3
Nvidia driver version: 570.86.15
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:                        52 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               52
On-line CPU(s) list:                  0-51
Vendor ID:                            GenuineIntel
Model name:                           Intel(R) Xeon(R) Platinum 8481C CPU @ 2.70GHz
CPU family:                           6
Model:                                143
Thread(s) per core:                   2
Core(s) per socket:                   26
Socket(s):                            1
Stepping:                             8
BogoMIPS:                             5399.99
Flags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology nonstop_tsc cpuid tsc_known_freq pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch ssbd ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rtm avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx_vnni avx512_bf16 arat avx512vbmi umip avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq rdpid cldemote movdiri movdir64b fsrm md_clear serialize tsxldtrk amx_bf16 avx512_fp16 amx_tile amx_int8 arch_capabilities
Hypervisor vendor:                    KVM
Virtualization type:                  full
L1d cache:                            1.2 MiB (26 instances)
L1i cache:                            832 KiB (26 instances)
L2 cache:                             52 MiB (26 instances)
L3 cache:                             105 MiB (1 instance)
NUMA node(s):                         1
NUMA node0 CPU(s):                    0-51
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; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:             Mitigation; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI SW loop, KVM SW loop
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] flashinfer-python==0.2.1.post2+cu124torch2.6
[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==26.3.0
[pip3] torch==2.6.0
[pip3] torchaudio==2.6.0
[pip3] torchvision==0.21.0
[pip3] transformers==4.49.0
[pip3] triton==3.2.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.8.0
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0	CPU Affinity	NUMA Affinity	GPU NUMA ID
GPU0	 X 	0-51	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=all
NVIDIA_REQUIRE_CUDA=cuda>=12.4 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 brand=tesla,driver>=535,driver<536 brand=unknown,driver>=535,driver<536 brand=nvidia,driver>=535,driver<536 brand=nvidiartx,driver>=535,driver<536 brand=geforce,driver>=535,driver<536 brand=geforcertx,driver>=535,driver<536 brand=quadro,driver>=535,driver<536 brand=quadrortx,driver>=535,driver<536 brand=titan,driver>=535,driver<536 brand=titanrtx,driver>=535,driver<536
NCCL_VERSION=2.20.5-1
NVIDIA_DRIVER_CAPABILITIES=compute,utility
NVIDIA_PRODUCT_NAME=CUDA
VLLM_USAGE_SOURCE=production-docker-image
CUDA_VERSION=12.4.0
VLLM_ALLOW_RUNTIME_LORA_UPDATING=true
LD_LIBRARY_PATH=/opt/venv/lib/python3.12/site-packages/cv2/../../lib64:/usr/local/nvidia/lib:/usr/local/nvidia/lib64
NCCL_CUMEM_ENABLE=0
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

Very long running tail requests aganist vllm report a p99 tail latency of 60s, when the real value is often much higher.

The request_latency_buckets for inference_time_request histogram cap at 60s, but there are many requests to a large language model that could take in significant excess of this. In practice folks use p90 and p99 on histograms to understand the latency buckets, but the top bucket limits how high the quantile estimation algorithm can go.

In general, the highest bucket should encompass the reasonable p99 distribution, and prometheus recommends:

Pick buckets suitable for the expected range of observed values

in https://prometheus.io/docs/practices/histograms/#quantiles

I would suggest adding at least 3 more exponential buckets - 120s, 240s, 480s - which means vllm out of the box would support p99 estimation for 8m and below. I have not heard of someone running requests longer than 8m, but it might be possible - so adding 960s (16m) would catch that.

I think it is reasonable to do that for all histogram metrics - this is a core golden signal (latency distribution) and so spending a slightly larger amount of space to track this is reasonable. For instance, Kubernetes regularly sees 500-1000 QPS read load on busy control planes, and the latency histogram for requests

Buckets: []float64{0.05, 0.1, 0.2, 0.4, 0.6, 0.8, 1.0, 1.25, 1.5, 2, 3,
				4, 5, 6, 8, 10, 15, 20, 30, 45, 60},

source

has 21 buckets. vLLM will see significantly longer requests, so even 32m might be feasible and would have minimal runtime cost.

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.

[smarterclayton]

One note, duplicated from #15202:

I'm not personally aware of anyone today running 16m or 32m request to a language model today. However, an 8B model with 128k context could in theory generate about 20 tokens per request per second (at a TPOT of 50ms at ~15 req/s on an H100). If there is only one very long output request running and a lot of prefill heavy requests you could imagine completing that in 6400s or 106m. So there is at least some data to suggest that these bucket sizes may even be too small today for today's models.

I think operationally the max runtime of an online serving request is likely to be less than 106m, but someone serving very long requests in stable environments could hit it.