vllm: [Usage]: Extremely slow inference with Llama 3.1 70b Instruct
Your current environment
Collecting environment information...
PyTorch version: 2.4.0+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
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 3.30.2
Libc version: glibc-2.35
Python version: 3.10.12 (main, Jul 29 2024, 16:56:48) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-6.5.0-1024-aws-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.1.105
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA A100-SXM4-40GB
GPU 1: NVIDIA A100-SXM4-40GB
GPU 2: NVIDIA A100-SXM4-40GB
GPU 3: NVIDIA A100-SXM4-40GB
GPU 4: NVIDIA A100-SXM4-40GB
GPU 5: NVIDIA A100-SXM4-40GB
GPU 6: NVIDIA A100-SXM4-40GB
GPU 7: NVIDIA A100-SXM4-40GB
Nvidia driver version: 535.183.01
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): 96
On-line CPU(s) list: 0-95
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) Platinum 8275CL CPU @ 3.00GHz
CPU family: 6
Model: 85
Thread(s) per core: 2
Core(s) per socket: 24
Socket(s): 2
Stepping: 7
BogoMIPS: 5999.99
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 arch_perfmon rep_good nopl xtopology nonstop_tsc cpuid aperfmperf 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 abm 3dnowprefetch invpcid_single pti fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid mpx avx512f avx512dq rdseed adx smap clflushopt clwb avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves ida arat pku ospke
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 1.5 MiB (48 instances)
L1i cache: 1.5 MiB (48 instances)
L2 cache: 48 MiB (48 instances)
L3 cache: 71.5 MiB (2 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-23,48-71
NUMA node1 CPU(s): 24-47,72-95
Vulnerability Gather data sampling: Unknown: Dependent on hypervisor status
Vulnerability Itlb multihit: KVM: Mitigation: VMX unsupported
Vulnerability L1tf: Mitigation; PTE Inversion
Vulnerability Mds: Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Meltdown: Mitigation; PTI
Vulnerability Mmio stale data: Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Retbleed: Vulnerable
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Vulnerable
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Retpolines; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Retpoline
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Versions of relevant libraries:
[pip3] numpy==1.26.4
[pip3] nvidia-nccl-cu12==2.20.5
[pip3] pyzmq==26.1.0
[pip3] torch==2.4.0
[pip3] torchvision==0.19.0
[pip3] transformers==4.44.0
[pip3] triton==3.0.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.5.4@4db5176d9758b720b05460c50ace3c01026eb158
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 GPU NUMA ID
GPU0 X NV12 NV12 NV12 NV12 NV12 NV12 NV12 0-23,48-71 0 N/A
GPU1 NV12 X NV12 NV12 NV12 NV12 NV12 NV12 0-23,48-71 0 N/A
GPU2 NV12 NV12 X NV12 NV12 NV12 NV12 NV12 0-23,48-71 0 N/A
GPU3 NV12 NV12 NV12 X NV12 NV12 NV12 NV12 0-23,48-71 0 N/A
GPU4 NV12 NV12 NV12 NV12 X NV12 NV12 NV12 24-47,72-95 1 N/A
GPU5 NV12 NV12 NV12 NV12 NV12 X NV12 NV12 24-47,72-95 1 N/A
GPU6 NV12 NV12 NV12 NV12 NV12 NV12 X NV12 24-47,72-95 1 N/A
GPU7 NV12 NV12 NV12 NV12 NV12 NV12 NV12 X 24-47,72-95 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
Following is the command I am using to run the Llama-3.1-70B-Instruct model
python -m vllm.entrypoints.openai.api_server --model meta-llama/Meta-Llama-3.1-70B-Instruct --tensor-parallel-size 8 --guided-decoding-backend lm-format-enforcer --enforce-eager.
The average generation throughput rarely exceeds 30 TPS. Do I need a better GPU or am I missing something else?
About this issue
- Original URL
- State: closed
- Created 2 years ago
- Comments: 23 (2 by maintainers)
Hi @yaronr i was review vllm config and i notice this config “–max-seq-len-to-capture” default is 8192, it will auto enable eager mode if your context len larger than this, i was change this config to 32k same as model len, and my result much faster. I see some config for parallel model and i will test soon
Hi @yaronr, is the 425 tokens/s including concurrent requests? Or is it the tokens/s for a single request? If it is for a single req, that would be pretty fast. The fastest I see for single req is ~30 tokens/s for Llama3.1 70b. If it is for 80 concurrent requests, does it translate to ~5 tokens/s per req? Thanks a lot! Followup – I confirmed on 8 * A100 that with TP=4, PP=2, the tokens/s is around 5. With TP=8, PP = 1, the tokens/s is around 30.
@MLlove0402 any updates? And if it improved could you please share your full deploy config command?
@iamhappytoo My figures were total output tokens (for all concurrent requests). It was designed to test total potential throughput, not necessarily at an acceptable per-request performance (because that changes by use case). From my experience, TP usually works better than PP, with few exceptions where PP + TP is better. It all boils down to how much chip-to-chip, card-to-card, and (hopefully not) node-to-node communication is required. The less cross-chip traffic, the better.