Why is the GPU KV cache usage very low?

[tammypi]

Your current environment

PyTorch version: 2.3.0+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.4 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0
Clang version: Could not collect
CMake version: version 3.26.3
Libc version: glibc-2.31

Python version: 3.9.12 (main, Apr  5 2022, 06:56:58)  [GCC 7.5.0] (64-bit runtime)
Python platform: Linux-5.10.0-182.0.0.95.oe2203sp3.x86_64-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: 12.2.91
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: NVIDIA L20
GPU 1: NVIDIA L20

Nvidia driver version: 535.154.05
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.2.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.2.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.2.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.2.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.2.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.2.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.2.0
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
架构：                              x86_64
CPU 运行模式：                      32-bit, 64-bit
字节序：                            Little Endian
Address sizes:                      46 bits physical, 57 bits virtual
CPU:                                104
在线 CPU 列表：                     0-103
每个核的线程数：                    2
每个座的核数：                      26
座：                                2
NUMA 节点：                         2
厂商 ID：                           GenuineIntel
CPU 系列：                          6
型号：                              106
型号名称：                          Intel(R) Xeon(R) Gold 5320 CPU @ 2.20GHz
步进：                              6
CPU MHz：                           2800.273
CPU 最大 MHz：                      3400.0000
CPU 最小 MHz：                      800.0000
BogoMIPS：                          4400.00
虚拟化：                            VT-x
L1d 缓存：                          2.4 MiB
L1i 缓存：                          1.6 MiB
L2 缓存：                           65 MiB
L3 缓存：                           78 MiB
NUMA 节点0 CPU：                    0-25,52-77
NUMA 节点1 CPU：                    26-51,78-103
Vulnerability Gather data sampling: Vulnerable: No microcode
Vulnerability Itlb multihit:        Not affected
Vulnerability L1tf:                 Not affected
Vulnerability Mds:                  Not affected
Vulnerability Meltdown:             Not affected
Vulnerability Mmio stale data:      Vulnerable: Clear CPU buffers attempted, no microcode; SMT vulnerable
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:           Vulnerable: eIBRS with unprivileged eBPF
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Not affected
标记：                              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 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 invpcid_single intel_ppin 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 wbnoinvd dtherm ida arat pln pts avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid fsrm md_clear pconfig flush_l1d arch_capabilities

Versions of relevant libraries:
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.22.3
[pip3] nvidia-nccl-cu11==2.14.3
[pip3] nvidia-nccl-cu12==2.20.5
[pip3] torch==2.3.0
[pip3] torchaudio==0.12.1
[pip3] torchinfo==1.7.2
[pip3] torchvision==0.13.1
[pip3] transformers==4.41.2
[pip3] transformers-stream-generator==0.0.4
[pip3] triton==2.3.0
[conda] blas                      1.0                         mkl  
[conda] cudatoolkit               11.3.1               h2bc3f7f_2  
[conda] ffmpeg                    4.3                  hf484d3e_0    pytorch
[conda] mkl                       2021.4.0           h06a4308_640  
[conda] mkl-service               2.4.0            py39h7f8727e_0  
[conda] mkl_fft                   1.3.1            py39hd3c417c_0  
[conda] mkl_random                1.2.2            py39h51133e4_0  
[conda] numpy                     1.22.3           py39he7a7128_0  
[conda] numpy-base                1.22.3           py39hf524024_0  
[conda] nvidia-nccl-cu11          2.14.3                   pypi_0    pypi
[conda] nvidia-nccl-cu12          2.20.5                   pypi_0    pypi
[conda] pytorch-mutex             1.0                        cuda    pytorch
[conda] torch                     2.3.0                    pypi_0    pypi
[conda] torchaudio                0.12.1               py39_cu113    pytorch
[conda] torchinfo                 1.7.2                    pypi_0    pypi
[conda] torchvision               0.13.1               py39_cu113    pytorch
[conda] transformers              4.41.2                   pypi_0    pypi
[conda] transformers-stream-generator 0.0.4                    pypi_0    pypi
[conda] triton                    2.3.0                    pypi_0    pypi
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.4.3
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    GPU1    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      SYS     0-25,52-77      0               N/A
GPU1    SYS      X      26-51,78-103    1               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

How would you like to use vllm

1.Model size: 14B
2.Performance: Average prompt throughput: 120.1 tokens/s, average generation throughput: 41.9 tokens/s, running: 1 request, swapped: 0 requests, pending: 0 requests, GPU KV cache usage: 1.0%, CPU KV cache usage: 0.0%.

The generation throughput speed is too slow. I think the reason might be the low GPU KV cache usage. How can I increase the GPU KV cache usage and improve the generation throughput speed?

[richardliaw]

If you're only running 1 request, your KV Cache is unlikely to be filled up. If you want to improve generation speed you can consider using FP8 or speculation

[umi-soft]

env

• 8 server，both has 2 T4 cards
• docker vllm-openai v0.11.2
• qwen3-32B float16
• tensor-parallel-size 8

The answer has not been output yet, and the log output is as follows：

Avg prompt throughput: 0.0 tokens/s, Avg generation throughputRunning: 8 reqs, Waiting:0 regs,GPU Ky cacheusage:0.5%，Prefix cache hit 68.6%

After about 3 minutes, the log output is as follows：

Avg prompt throughput: 0.0 tokens/s, Avg generation throughputRunning: 8 reqs, Waiting:0 regs,GPU Ky cacheusage:4.5%，Prefix cache hit 68.6%

@richardliaw @tammypi

[liut89]

@tammypi 你好请问你的GPU KV cache usage是怎么统计的？我在本地跑 benchmark_throughput 想查看kvcache使用情况
@tammypi How to get the GPU KV cache usage? I'm running benchmark_throughput locally and want to check the KV cache usage."

[jaysontree]

It is displayed at INFO level in VLLM log.
vllm打印的日志里就有。