[Bug]: Issue of Unstable Output for Identical Queries

[skylee-01]

Your current environment

INFO 06-10 00:07:35 [init.py:239] 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 22.04.3 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.31.4
Libc version: glibc-2.35

Python version: 3.12.8 (main, Dec 4 2024, 08:54:12) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.4.54-1.0.0.std7c.el7.2.x86_64-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 GeForce RTX 4090
GPU 1: NVIDIA GeForce RTX 4090
GPU 2: NVIDIA GeForce RTX 4090
GPU 3: NVIDIA GeForce RTX 4090
GPU 4: NVIDIA GeForce RTX 4090
GPU 5: NVIDIA GeForce RTX 4090
GPU 6: NVIDIA GeForce RTX 4090
GPU 7: NVIDIA GeForce RTX 4090

Nvidia driver version: 535.104.05
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): 176
On-line CPU(s) list: 0-175
Vendor ID: GenuineIntel
BIOS Vendor ID: Intel(R) Corporation
Model name: Intel(R) Xeon(R) Platinum 8458P
BIOS Model name: Intel(R) Xeon(R) Platinum 8458P
CPU family: 6
Model: 143
Thread(s) per core: 2
Core(s) per socket: 44
Socket(s): 2
Stepping: 8
Frequency boost: enabled
CPU max MHz: 2701.0000
CPU min MHz: 800.0000
BogoMIPS: 5400.00
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 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 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 rdpid cldemote movdiri movdir64b md_clear pconfig flush_l1d arch_capabilities
Virtualization: VT-x
L1d cache: 4.1 MiB (88 instances)
L1i cache: 2.8 MiB (88 instances)
L2 cache: 176 MiB (88 instances)
L3 cache: 165 MiB (2 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-43,88-131
NUMA node1 CPU(s): 44-87,132-175
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: 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 IBRS, IBPB conditional, RSB filling
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected

Versions of relevant libraries:
[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-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==26.2.1
[pip3] torch==2.6.0
[pip3] torchaudio==2.6.0
[pip3] torchvision==0.21.0
[pip3] transformers==4.52.4
[pip3] triton==3.2.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.8.5
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 NIC0 NIC1 CPU Affinity NUMA Affinity GPU NUMA ID
GPU0 X NODE NODE NODE SYS SYS SYS SYS NODE NODE 0-43,88-131 0 N/A
GPU1 NODE X NODE NODE SYS SYS SYS SYS NODE NODE 0-43,88-131 0 N/A
GPU2 NODE NODE X NODE SYS SYS SYS SYS NODE NODE 0-43,88-131 0 N/A
GPU3 NODE NODE NODE X SYS SYS SYS SYS NODE NODE 0-43,88-131 0 N/A
GPU4 SYS SYS SYS SYS X NODE NODE NODE SYS SYS 44-87,132-175 1 N/A
GPU5 SYS SYS SYS SYS NODE X NODE NODE SYS SYS 44-87,132-175 1 N/A
GPU6 SYS SYS SYS SYS NODE NODE X NODE SYS SYS 44-87,132-175 1 N/A
GPU7 SYS SYS SYS SYS NODE NODE NODE X SYS SYS 44-87,132-175 1 N/A
NIC0 NODE NODE NODE NODE SYS SYS SYS SYS X PIX
NIC1 NODE NODE NODE NODE SYS SYS SYS SYS PIX 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
NIC1: mlx5_1

NVIDIA_VISIBLE_DEVICES=all
NVIDIA_REQUIRE_CUDA=cuda>=12.1 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
NCCL_VERSION=2.17.1-1
NVIDIA_DRIVER_CAPABILITIES=compute,utility
NVIDIA_PRODUCT_NAME=CUDA
VLLM_USAGE_SOURCE=production-docker-image
NVIDIA_CUDA_END_OF_LIFE=1
CUDA_VERSION=12.1.0
LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64
NCCL_CUMEM_ENABLE=0
PYTORCH_NVML_BASED_CUDA_CHECK=1
TORCHINDUCTOR_COMPILE_THREADS=1
CUDA_MODULE_LOADING=LAZY

🐛 Describe the bug

When the enforce_eager parameter is set to False, an anomalous behavior was observed where the service intermittently produced incorrect results for the initial query in certain operational contexts.

Problem Description

• In serial request scenarios involving single-data inputs, regardless of the total number of requests issued, the first request consistently returned erroneous outputs (refer to Figure for visual representation).
• When executing concurrent requests with single-data payloads, instances of incorrect outputs occurred randomly, with a significantly elevated frequency compared to serial execution.
• Restoration of normal functionality was observed for both serial and concurrent request scenarios when the enforce_eager parameter was set to True.
• This phenomenon was reproducibly observed across both online service deployments and offline engine testing environments.

Problem Analysis
A working hypothesis suggests that the instability may be attributed to the CUDA Graph's computational behavior during the initial capture of tensors with specific geometric configurations.

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.

[skylee-01]

[skylee-01]

# -*- coding: utf-8 -*-
import sys
import os

os.environ["CUDA_VISIBLE_DEVICES"] = "0"
os.environ["VLLM_USE_V1"] = "1"

custom_vllm_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), 'vllm')

if custom_vllm_path not in sys.path:
   sys.path.insert(0, custom_vllm_path)

print(sys.path)

from vllm import LLM, SamplingParams

prompts = [
    "Please generate an appropriate response based on the following context:\n\nContext:\nThis is a sample conversation scenario\n\nUser Input:\nHello, please help me complete a simple task",
]

sampling_params = SamplingParams(temperature=0, max_tokens=200)


model = os.getenv("MODEL_PATH", "path/to/your/model")


llm = LLM(model=model,
            # enforce_eager=True, 
            max_model_len=100,
            gpu_memory_utilization=0.7,
        )

for i in range(5):
    print("************************************************")
    outputs = llm.generate(prompts, sampling_params)
    for output in outputs:
        answer = output.outputs[0].text
        print(f"answer:{answer}")

[skylee-01]

@youkaichao Please take a look when you have a moment.

[skylee-01]

Synchronize the latest analysis
Adding the configuration compilation_config=CompilationConfig(custom_ops=["+rms_norm"]) restored normal output.
How can I further troubleshoot this?

[princepride]

Hmm, I use the "Qwen/Qwen2.5-3B-Instruct" model but can't reproduce your bug, can you tell me which model you use?

[skylee-01]

Hmm, I use the "Qwen/Qwen2.5-3B-Instruct" model but can't reproduce your bug, can you tell me which model you use?

Thank you for your reply. We are using the model fine-tuned on Qwen/Qwen2.5-3B-Instruct.
Let me sync my latest analysis:
I compared the compiled files generated with and without the configuration `compilation_config=CompilationConfig(custom_ops=["+rms_norm"]). It was found that the problematic file contains an additional Triton operator generated by the compiler. I suspect this operator is causing the issue.

[princepride]

INFO 06-10 12:50:33 [core.py:171] init engine (profile, create kv cache, warmup model) took 33.36 seconds
************************************************
Adding requests: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 297.47it/s]
Processed prompts: 100%|████████████████████████████████████████████████████████████| 1/1 [00:00<00:00,  1.40it/s, est. speed input: 44.76 toks/s, output: 95.10 toks/s]
answer:.

Assistant:
Of course! I'd be happy to help you with a simple task. What task would you like assistance with? Please provide more details so I can assist you better.  How can I assist you today?  Let me know the specifics of the task you need help with.  Is there a particular area or subject you
************************************************
Adding requests: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1374.73it/s]
Processed prompts: 100%|████████████████████████████████████████████████████████████| 1/1 [00:00<00:00,  1.42it/s, est. speed input: 45.47 toks/s, output: 96.62 toks/s]
answer:.

Assistant:
Of course! I'd be happy to help you with a simple task. What task would you like assistance with? Please provide more details so I can assist you better.  How can I assist you today?  Let me know the specifics of the task you need help with.  Is there a particular area or subject you
************************************************
Adding requests: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 931.45it/s]
Processed prompts: 100%|████████████████████████████████████████████████████████████| 1/1 [00:00<00:00,  1.42it/s, est. speed input: 45.47 toks/s, output: 96.62 toks/s]
answer:.

Assistant:
Of course! I'd be happy to help you with a simple task. What task would you like assistance with? Please provide more details so I can assist you better.  How can I assist you today?  Let me know the specifics of the task you need help with.  Is there a particular area or subject you
************************************************
Adding requests: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1823.61it/s]
Processed prompts: 100%|████████████████████████████████████████████████████████████| 1/1 [00:00<00:00,  1.42it/s, est. speed input: 45.38 toks/s, output: 96.43 toks/s]
answer:.

Assistant:
Of course! I'd be happy to help you with a simple task. What task would you like assistance with? Please provide more details so I can assist you better.  How can I assist you today?  Let me know the specifics of the task you need help with.  Is there a particular area or subject you
************************************************
Adding requests: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1392.53it/s]
Processed prompts: 100%|████████████████████████████████████████████████████████████| 1/1 [00:00<00:00,  1.42it/s, est. speed input: 45.40 toks/s, output: 96.47 toks/s]
answer:.

Assistant:
Of course! I'd be happy to help you with a simple task. What task would you like assistance with? Please provide more details so I can assist you better.  How can I assist you today?  Let me know the specifics of the task you need help with.  Is there a particular area or subject you

That's weird, I followed your example code and run the original Qwen/Qwen2.5-3B-Instruct model, seems the output is OK.

Is my script problem?

from vllm import LLM, SamplingParams

prompts = [
    "Please generate an appropriate response based on the following context:\n\nContext:\nThis is a sample conversation scenario\n\nUser Input:\nHello, please help me complete a simple task",
]

sampling_params = SamplingParams(temperature=0, max_tokens=200)


model = "Qwen/Qwen2.5-3B-Instruct"


llm = LLM(model=model,
            # enforce_eager=True, 
            max_model_len=100,
            gpu_memory_utilization=0.7,
        )

for i in range(5):
    print("************************************************")
    outputs = llm.generate(prompts, sampling_params)
    for output in outputs:
        answer = output.outputs[0].text
        print(f"answer:{answer}")

[skylee-01]

INFO 06-10 12:50:33 [core.py:171] init engine (profile, create kv cache, warmup model) took 33.36 seconds
************************************************
Adding requests: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 297.47it/s]
Processed prompts: 100%|████████████████████████████████████████████████████████████| 1/1 [00:00<00:00,  1.40it/s, est. speed input: 44.76 toks/s, output: 95.10 toks/s]
answer:.

Assistant:
Of course! I'd be happy to help you with a simple task. What task would you like assistance with? Please provide more details so I can assist you better.  How can I assist you today?  Let me know the specifics of the task you need help with.  Is there a particular area or subject you
************************************************
Adding requests: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1374.73it/s]
Processed prompts: 100%|████████████████████████████████████████████████████████████| 1/1 [00:00<00:00,  1.42it/s, est. speed input: 45.47 toks/s, output: 96.62 toks/s]
answer:.

Assistant:
Of course! I'd be happy to help you with a simple task. What task would you like assistance with? Please provide more details so I can assist you better.  How can I assist you today?  Let me know the specifics of the task you need help with.  Is there a particular area or subject you
************************************************
Adding requests: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 931.45it/s]
Processed prompts: 100%|████████████████████████████████████████████████████████████| 1/1 [00:00<00:00,  1.42it/s, est. speed input: 45.47 toks/s, output: 96.62 toks/s]
answer:.

Assistant:
Of course! I'd be happy to help you with a simple task. What task would you like assistance with? Please provide more details so I can assist you better.  How can I assist you today?  Let me know the specifics of the task you need help with.  Is there a particular area or subject you
************************************************
Adding requests: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1823.61it/s]
Processed prompts: 100%|████████████████████████████████████████████████████████████| 1/1 [00:00<00:00,  1.42it/s, est. speed input: 45.38 toks/s, output: 96.43 toks/s]
answer:.

Assistant:
Of course! I'd be happy to help you with a simple task. What task would you like assistance with? Please provide more details so I can assist you better.  How can I assist you today?  Let me know the specifics of the task you need help with.  Is there a particular area or subject you
************************************************
Adding requests: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 1392.53it/s]
Processed prompts: 100%|████████████████████████████████████████████████████████████| 1/1 [00:00<00:00,  1.42it/s, est. speed input: 45.40 toks/s, output: 96.47 toks/s]
answer:.

Assistant:
Of course! I'd be happy to help you with a simple task. What task would you like assistance with? Please provide more details so I can assist you better.  How can I assist you today?  Let me know the specifics of the task you need help with.  Is there a particular area or subject you

That's weird, I followed your example code and run the original Qwen/Qwen2.5-3B-Instruct model, seems the output is OK.

Is my script problem?

from vllm import LLM, SamplingParams

prompts = [
    "Please generate an appropriate response based on the following context:\n\nContext:\nThis is a sample conversation scenario\n\nUser Input:\nHello, please help me complete a simple task",
]

sampling_params = SamplingParams(temperature=0, max_tokens=200)


model = "Qwen/Qwen2.5-3B-Instruct"


llm = LLM(model=model,
            # enforce_eager=True, 
            max_model_len=100,
            gpu_memory_utilization=0.7,
        )

for i in range(5):
    print("************************************************")
    outputs = llm.generate(prompts, sampling_params)
    for output in outputs:
        answer = output.outputs[0].text
        print(f"answer:{answer}")

Your script has no issues; this case occurs sporadically. Our service handles a large amount of traffic, and such cases have been identified from monitoring data.

[princepride]

Got it, seems exceed my ability, if you can't contact youkaichao, you can contact other maintainer

[princepride]

By the way, does v0 engine occurred same bug?

[youkaichao]

cc @zou3519 , it seems to be caused by inductor's triton kernel.