[Bug]: vLLM outputs are not reproducible

[FelixSchneiderZoom]

Your current environment

The output of python collect_env.py

Collecting environment information...
==============================
        System Info
==============================
OS                           : Ubuntu 22.04.5 LTS (x86_64)
GCC version                  : (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version                : Could not collect
CMake version                : version 3.22.1
Libc version                 : glibc-2.35

==============================
       PyTorch Info
==============================
PyTorch version              : 2.7.0+cu126
Is debug build               : False
CUDA used to build PyTorch   : 12.6
ROCM used to build PyTorch   : N/A

==============================
      Python Environment
==============================
Python version               : 3.11.11 (main, Dec 11 2024, 16:28:39) [GCC 11.2.0] (64-bit runtime)
Python platform              : Linux-5.15.0-135-generic-x86_64-with-glibc2.35

==============================
       CUDA / GPU Info
==============================
Is CUDA available            : True
CUDA runtime version         : Could not collect
CUDA_MODULE_LOADING set to   : LAZY
GPU models and configuration :
GPU 0: NVIDIA H100 80GB HBM3
GPU 1: NVIDIA H100 80GB HBM3
GPU 2: NVIDIA H100 80GB HBM3
GPU 3: NVIDIA H100 80GB HBM3
GPU 4: NVIDIA H100 80GB HBM3
GPU 5: NVIDIA H100 80GB HBM3
GPU 6: NVIDIA H100 80GB HBM3
GPU 7: NVIDIA H100 80GB HBM3

Nvidia driver version        : 550.127.05
cuDNN version                : Could not collect
HIP runtime version          : N/A
MIOpen runtime version       : N/A
Is XNNPACK available         : True

==============================
          CPU Info
==============================
Architecture:                         x86_64
CPU op-mode(s):                       32-bit, 64-bit
Address sizes:                        46 bits physical, 57 bits virtual
Byte Order:                           Little Endian
CPU(s):                               224
On-line CPU(s) list:                  0-223
Vendor ID:                            GenuineIntel
Model name:                           Intel(R) Xeon(R) Platinum 8480+
CPU family:                           6
Model:                                143
Thread(s) per core:                   2
Core(s) per socket:                   56
Socket(s):                            2
Stepping:                             8
BogoMIPS:                             4000.00
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 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 cat_l2 cdp_l3 invpcid_single cdp_l2 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 avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization:                       VT-x
L1d cache:                            5.3 MiB (112 instances)
L1i cache:                            3.5 MiB (112 instances)
L2 cache:                             224 MiB (112 instances)
L3 cache:                             210 MiB (2 instances)
NUMA node(s):                         2
NUMA node0 CPU(s):                    0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62,64,66,68,70,72,74,76,78,80,82,84,86,88,90,92,94,96,98,100,102,104,106,108,110,112,114,116,118,120,122,124,126,128,130,132,134,136,138,140,142,144,146,148,150,152,154,156,158,160,162,164,166,168,170,172,174,176,178,180,182,184,186,188,190,192,194,196,198,200,202,204,206,208,210,212,214,216,218,220,222
NUMA node1 CPU(s):                    1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63,65,67,69,71,73,75,77,79,81,83,85,87,89,91,93,95,97,99,101,103,105,107,109,111,113,115,117,119,121,123,125,127,129,131,133,135,137,139,141,143,145,147,149,151,153,155,157,159,161,163,165,167,169,171,173,175,177,179,181,183,185,187,189,191,193,195,197,199,201,203,205,207,209,211,213,215,217,219,221,223
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 and seccomp
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==12.6.4.1
[pip3] nvidia-cuda-cupti-cu12==12.6.80
[pip3] nvidia-cuda-nvrtc-cu12==12.6.77
[pip3] nvidia-cuda-runtime-cu12==12.6.77
[pip3] nvidia-cudnn-cu12==9.5.1.17
[pip3] nvidia-cufft-cu12==11.3.0.4
[pip3] nvidia-cufile-cu12==1.11.1.6
[pip3] nvidia-curand-cu12==10.3.7.77
[pip3] nvidia-cusolver-cu12==11.7.1.2
[pip3] nvidia-cusparse-cu12==12.5.4.2
[pip3] nvidia-cusparselt-cu12==0.6.3
[pip3] nvidia-htop==1.2.0
[pip3] nvidia-nccl-cu12==2.26.2
[pip3] nvidia-nvjitlink-cu12==12.6.85
[pip3] nvidia-nvtx-cu12==12.6.77
[pip3] pytorch-lightning==2.5.1.post0
[pip3] pyzmq==26.4.0
[pip3] torch==2.7.0
[pip3] torchaudio==2.7.0
[pip3] torchmetrics==0.10.3
[pip3] torchvision==0.22.0
[pip3] transformers==4.52.4
[pip3] triton==3.3.0
[conda] numpy                     1.26.4                   pypi_0    pypi
[conda] nvidia-cublas-cu12        12.6.4.1                 pypi_0    pypi
[conda] nvidia-cuda-cupti-cu12    12.6.80                  pypi_0    pypi
[conda] nvidia-cuda-nvrtc-cu12    12.6.77                  pypi_0    pypi
[conda] nvidia-cuda-runtime-cu12  12.6.77                  pypi_0    pypi
[conda] nvidia-cudnn-cu12         9.5.1.17                 pypi_0    pypi
[conda] nvidia-cufft-cu12         11.3.0.4                 pypi_0    pypi
[conda] nvidia-cufile-cu12        1.11.1.6                 pypi_0    pypi
[conda] nvidia-curand-cu12        10.3.7.77                pypi_0    pypi
[conda] nvidia-cusolver-cu12      11.7.1.2                 pypi_0    pypi
[conda] nvidia-cusparse-cu12      12.5.4.2                 pypi_0    pypi
[conda] nvidia-cusparselt-cu12    0.6.3                    pypi_0    pypi
[conda] nvidia-htop               1.2.0                    pypi_0    pypi
[conda] nvidia-nccl-cu12          2.26.2                   pypi_0    pypi
[conda] nvidia-nvjitlink-cu12     12.6.85                  pypi_0    pypi
[conda] nvidia-nvtx-cu12          12.6.77                  pypi_0    pypi
[conda] pytorch-lightning         2.5.1.post0              pypi_0    pypi
[conda] pyzmq                     26.4.0                   pypi_0    pypi
[conda] torch                     2.7.0                    pypi_0    pypi
[conda] torchaudio                2.7.0                    pypi_0    pypi
[conda] torchmetrics              0.10.3                   pypi_0    pypi
[conda] torchvision               0.22.0                   pypi_0    pypi
[conda] transformers              4.52.4                   pypi_0    pypi
[conda] triton                    3.3.0                    pypi_0    pypi

==============================
         vLLM Info
==============================
ROCM Version                 : Could not collect
Neuron SDK Version           : N/A
vLLM Version                 : 0.9.1
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      NV18    NV18    NV18    NV18    NV18    NV18    NV18    0,2,4,6,8,10    0               N/A
GPU1    NV18     X      NV18    NV18    NV18    NV18    NV18    NV18    0,2,4,6,8,10    0               N/A
GPU2    NV18    NV18     X      NV18    NV18    NV18    NV18    NV18    0,2,4,6,8,10    0               N/A
GPU3    NV18    NV18    NV18     X      NV18    NV18    NV18    NV18    0,2,4,6,8,10    0               N/A
GPU4    NV18    NV18    NV18    NV18     X      NV18    NV18    NV18    1,3,5,7,9,11    1               N/A
GPU5    NV18    NV18    NV18    NV18    NV18     X      NV18    NV18    1,3,5,7,9,11    1               N/A
GPU6    NV18    NV18    NV18    NV18    NV18    NV18     X      NV18    1,3,5,7,9,11    1               N/A
GPU7    NV18    NV18    NV18    NV18    NV18    NV18    NV18     X      1,3,5,7,9,11    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

==============================
     Environment Variables
==============================
CUDA_VISIBLE_DEVICES=3
CUDA_VISIBLE_DEVICES=3
VLLM_CONFIGURE_LOGGING=0
VLLM_ENABLE_V1_MULTIPROCESSING=0
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

When using AsyncLLM, the outputs are not reproducible, even when following the instructions in the documentation (https://docs.vllm.ai/en/stable/usage/reproducibility.html).

For example, running the below script multiple times and diff-ing the outputs reveals differences in the logprobs even for outputs that are identical (which shouldn't be the case) and in some cases changed outputs. It seems like it's necessary to make a larger number of requests at the same time to observe the effect (possibly it's related to the scheduling, so a certain pressure is required).

When setting temperature=0, I expect totally deterministic output from the model, and this is extremely important to our use cases, where we might need to debug problematic model outputs for particular inputs. For this small Qwen2 model, the effect is relatively minor, but for larger models I observed much larger differences.

import asyncio
import itertools
from hashlib import sha1

import datasets
import vllm
from vllm.v1.engine.async_llm import AsyncLLM


async def main():
    # 198 prompts, actual contents don't matter
    prompts = datasets.load_dataset("cais/mmlu", "high_school_geography")["test"]["question"]
    llm = AsyncLLM.from_engine_args(vllm.AsyncEngineArgs(model="Qwen/Qwen2-0.5B-Instruct", disable_log_requests=True))

    results = await asyncio.gather(*(request(llm, prompt) for prompt in prompts))
    for result in results:
        print(
            f"{result.request_id}: {result.outputs[0].cumulative_logprob} {result.outputs[0].text[:40]} {sha1(result.outputs[0].text.encode('utf-8')).hexdigest()}"
        )


request_counter = itertools.count()


async def request(llm, prompt: str):
    request_id = next(request_counter)
    chat_prompt = [{"role": "user", "content": prompt}]
    tokens = llm.tokenizer.tokenizer.apply_chat_template(chat_prompt, add_generation_prompt=True, tokenize=True)
    results_generator = llm.generate(
        vllm.TokensPrompt(prompt_token_ids=tokens),
        sampling_params=vllm.SamplingParams(temperature=0.0, max_tokens=1024, logprobs=0),
        request_id=str(request_id),
    )
    final_output = None
    async for request_output in results_generator:
        final_output = request_output
    return final_output


if __name__ == "__main__":
    asyncio.run(main())

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.

[princepride]

Hmm, seems the same reason cause this bug: #19403

[FelixSchneiderZoom]

Update: I also reproduced this bug with the LLM API:

from hashlib import sha1

import datasets
import vllm


def main():
    # 198 prompts, actual contents don't matter
    prompts = datasets.load_dataset("cais/mmlu", "high_school_geography")["test"]["question"]
    llm = vllm.LLM("Qwen/Qwen2-0.5B-Instruct")

    results = llm.chat(
        [[{"role": "user", "content": prompt}] for prompt in prompts],
        sampling_params=vllm.SamplingParams(temperature=0.0, max_tokens=1024, logprobs=0),
    )
    for i, result in enumerate(results):
        print(
            f"{i}: {result.outputs[0].cumulative_logprob} {result.outputs[0].text[:40]} {sha1(result.outputs[0].text.encode('utf-8')).hexdigest()}"
        )


if __name__ == "__main__":
    main()

[FelixSchneiderZoom]

Using compilation_config=CompilationConfig(custom_ops=["+rms_norm"]), I was able to restore deterministic behavior for LLM, but not for AsyncLLM.

[FelixSchneiderZoom]

Setting enforce_eager=True restores deterministic behavior for AsyncLLM, but probably loses a lot of the benefits of vLLM. I'm not deep enough in the custom ops to debug this further. I hope we can get determinism in both APIs.

[zou3519]

@FelixSchneiderZoom could you try torch.use_deterministic_algorithms(True) (add it to somewhere at the top of your file) please?

[zou3519]

This repros pretty consistently for me (at least one of the three hashes is different), even with use_deterministic_algorithms:

from hashlib import sha1

import datasets
import vllm

import torch
torch.use_deterministic_algorithms(True)


def main():
    # 198 prompts, actual contents don't matter
    prompts = datasets.load_dataset("cais/mmlu", "high_school_geography")["test"]["question"]
    # llm = vllm.LLM("Qwen/Qwen2-0.5B-Instruct", compilation_config={"custom_ops": ["+rms_norm"]})
    # llm = vllm.LLM("Qwen/Qwen2-0.5B-Instruct", enforce_eager=True)
    llm = vllm.LLM("Qwen/Qwen2-0.5B-Instruct", compilation_config={"use_cudagraph": False})  # different hash

    def do_once():
        results = llm.chat(
            [[{"role": "user", "content": prompt}] for prompt in prompts],
            sampling_params=vllm.SamplingParams(temperature=0.0, max_tokens=1024, logprobs=0),
        )
        outputs = "".join([r.outputs[0].text for r in results])
        outputs = outputs.encode('utf-8')
        print("=" * 80)
        print(sha1(outputs).hexdigest())
        print("=" * 80)

    do_once()
    do_once()
    do_once()

if __name__ == "__main__":
    main()

[zou3519]

@FelixSchneiderZoom setting VLLM_ENABLE_V1_MULTIPROCESSING=0 seems to make the problem go away for me for the LLM case

[BoyuanFeng]

yes, VLLM_ENABLE_V1_MULTIPROCESSING=0 CUBLAS_WORKSPACE_CONFIG=:4096:8 python3 repro.py gives deterministic results: https://pastebin.com/7aBnjhJ8

Still investigating why VLLM_ENABLE_V1_MULTIPROCESSING or rms_norm op leads to the issue.

[FelixSchneiderZoom]

@zou3519 I mainly care about the AsyncLLM case.
torch.use_deterministic_algorithms does not work for me, but the error message mentions CUBLAS_WORKSPACE_CONFIG. Unfortunately, for AsyncLLM, this does not appear to be enough.

$ VLLM_ENABLE_V1_MULTIPROCESSING=0 CUBLAS_WORKSPACE_CONFIG=:4096:8 python test_async.py > a.txt
$ VLLM_ENABLE_V1_MULTIPROCESSING=0 CUBLAS_WORKSPACE_CONFIG=:4096:8 python test_async.py > b.txt
$ diff a.txt b.txt
(non-empty output)

(test_async.py is the script from my initial post)

It does seem to work for the LLM case, but it's not trivial to restructure our code around this, and there is no guarantee that the potentially different ordering will not break LLM as well.

[zou3519]

Thanks for the datapoint, we'll investigate more.