[Bug]: TTFT Performance Regression in vLLM v0.7.0 Compared to v0.6.1.post2

[asleepykitty]

Your current environment

The output of `python collect_env.py` on vLLM 0.7.0

INFO 03-14 21:24:53 __init__.py:183] Automatically detected platform cuda.
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: CBL-Mariner/Linux (x86_64)
GCC version: (GCC) 11.2.0
Clang version: Could not collect
CMake version: version 3.21.4
Libc version: glibc-2.35

Python version: 3.10.14 (main, Jul 14 2024, 22:24:12) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-1070-azure-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 11.8.89
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA A100 80GB PCIe
GPU 1: NVIDIA A100 80GB PCIe
GPU 2: NVIDIA A100 80GB PCIe
GPU 3: NVIDIA A100 80GB PCIe

Nvidia driver version: 550.90.07
cuDNN version: Probably one of the following:
/usr/lib/libcudnn.so.8.9.5
/usr/lib/libcudnn_adv_infer.so.8.9.5
/usr/lib/libcudnn_adv_train.so.8.9.5
/usr/lib/libcudnn_cnn_infer.so.8.9.5
/usr/lib/libcudnn_cnn_train.so.8.9.5
/usr/lib/libcudnn_ops_infer.so.8.9.5
/usr/lib/libcudnn_ops_train.so.8.9.5
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):                               96
On-line CPU(s) list:                  0-95
Vendor ID:                            AuthenticAMD
Model name:                           AMD EPYC 7V13 64-Core Processor
CPU family:                           25
Model:                                1
Thread(s) per core:                   1
Core(s) per socket:                   48
Socket(s):                            2
Stepping:                             1
BogoMIPS:                             4890.88
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 tsc_reliable nonstop_tsc cpuid extd_apicid aperfmperf pni pclmulqdq ssse3 fma cx16 pcid sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw topoext perfctr_core invpcid_single vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves clzero xsaveerptr rdpru arat umip vaes vpclmulqdq rdpid fsrm
Hypervisor vendor:                    Microsoft
Virtualization type:                  full
L1d cache:                            3 MiB (96 instances)
L1i cache:                            3 MiB (96 instances)
L2 cache:                             48 MiB (96 instances)
L3 cache:                             384 MiB (12 instances)
NUMA node(s):                         4
NUMA node0 CPU(s):                    0-23
NUMA node1 CPU(s):                    24-47
NUMA node2 CPU(s):                    48-71
NUMA node3 CPU(s):                    72-95
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:   Mitigation; safe RET, no microcode
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 Not affected
Vulnerability Srbds:                  Not affected
Vulnerability Tsx async abort:        Not affected

Versions of relevant libraries:
[pip3] flake8==4.0.1.1
[pip3] mypy==1.11.2
[pip3] mypy-extensions==1.0.0
[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==12.570.86
[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] sentence-transformers==3.0.0
[pip3] torch==2.5.1+cu124
[pip3] torchvision==0.20.1+cu124
[pip3] transformers==4.49.0
[pip3] triton==3.1.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.7.0
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    GPU1    GPU2    GPU3    NIC0    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      NV12    SYS     SYS     NODE    0-23    0               N/A
GPU1    NV12     X      SYS     SYS     SYS     24-47   1               N/A
GPU2    SYS     SYS      X      NV12    SYS     48-71   2               N/A
GPU3    SYS     SYS     NV12     X      SYS     72-95   3               N/A
NIC0    NODE    SYS     SYS     SYS      X

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

NIC Legend:

  NIC0: mlx5_0

NVIDIA_VISIBLE_DEVICES=1
CUDA_VISIBLE_DEVICES=1
CUDA_VISIBLE_DEVICES=1
LD_LIBRARY_PATH=/home/coder/proxima-hosted-language-models/build/proxima-hosted-language-models/environments/development-venv/lib/python3.10/site-packages/cv2/../../lib64:/usr/local/cuda/lib64:/usr/local/cuda/compat/:/usr/local/cuda/targets/x86_64-linux/lib/:/export/apps/hadoop/latest/lib/native/:/usr/local/cuda/lib64:/usr/local/cuda/compat/:/usr/local/cuda/targets/x86_64-linux/lib/:/export/apps/hadoop/latest/lib/native/:/usr/local/cuda/lib64:/usr/local/cuda/compat/:/usr/local/cuda/targets/x86_64-linux/lib/:/export/apps/hadoop/latest/lib/native/:/usr/local/cuda/lib64:/usr/local/cuda/compat/:/usr/local/cuda/targets/x86_64-linux/lib/:/export/apps/hadoop/latest/lib/native/:
NCCL_CUMEM_ENABLE=0
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

Some changes on vLLM 0.6.1.post2

vllm==0.6.1.post2
torch==2.4.0+cu121
torchvision==0.19.0
vllm-flash-attn==2.6.1
xformers==0.0.27.post2
transformers==4.46.0
pydantic==2.9.0

🐛 Describe the bug

Summary
After upgrading from vLLM 0.6.1.post2 to vLLM 0.7.0, we observed a significant increase in TTFT. While e2e latency and throughput improved in v0.7.0, TTFT became 3x slower.

We use:

• Our own model. The base model is Llama3 8B
• QPS: 1.2
• TTFT (v0.7.0): 0.39s
• TTFT (v0.6.1.post2): 0.13s
• Input Prompt: 4K tokens
• Output Tokens: 100

Reproduction Steps

1. Get TTFT on v0.7.0

• Run vLLM server

vllm serve /export/content/data/tmp/custom_model_path/proxima/model_resources \
    --tokenizer /export/content/data/tmp/custom_model_path/proxima/tokenizer_resources \
    --tensor-parallel-size 1 \
    --max-num-batched-tokens 2048 \
    --gpu-memory-utilization 0.9 \
    --enable-chunked-prefill \
    --use-v2-block-manager \
    --trust-remote-code \
    --guided-decoding-backend outlines

• Send requests to vLLM backend

python benchmark_serving.py --backend vllm --model (our model) --request-rate 1.2 --save-result

2. Get TTFT on v0.6.1.post2

• Run vLLM server

vllm serve /export/content/data/tmp/custom_model_path/proxima/model_resources \
    --tokenizer /export/content/data/tmp/custom_model_path/proxima/tokenizer_resources \
    --tensor-parallel-size 1 \
    --gpu-memory-utilization 0.9 \
    --enable-chunked-prefill \
    --use-v2-block-manager \
    --trust-remote-code \
    --guided-decoding-backend outlines

• Send requests to vLLM backend

python benchmark_serving.py --backend vllm --model (our model) --request-rate 1.2 --save-result

3. Compare TTFT results

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.

[russellb]

There have been multiple releases since v0.7.0. Can you try the latest from main ? 0.8.0 is about to come out any day now.