vllm: [Bug]: When enabling LoRA, greedy search got different answers.

Your current environment

The output of `python collect_env.py`
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 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.31

Python version: 3.10.14 (main, Apr  6 2024, 18:45:05) [GCC 9.4.0] (64-bit runtime)
Python platform: Linux-5.15.0-25-generic-x86_64-with-glibc2.31
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

Nvidia driver version: 535.86.10
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
Byte Order:                      Little Endian
Address sizes:                   52 bits physical, 57 bits virtual
CPU(s):                          96
On-line CPU(s) list:             0-95
Thread(s) per core:              1
Core(s) per socket:              48
Socket(s):                       2
NUMA node(s):                    8
Vendor ID:                       GenuineIntel
CPU family:                      6
Model:                           143
Model name:                      Intel(R) Xeon(R) Platinum 8468
Stepping:                        8
CPU MHz:                         2100.000
CPU max MHz:                     2100.0000
CPU min MHz:                     800.0000
BogoMIPS:                        4200.00
L1d cache:                       4.5 MiB
L1i cache:                       3 MiB
L2 cache:                        192 MiB
L3 cache:                        210 MiB
NUMA node0 CPU(s):               0-11
NUMA node1 CPU(s):               12-23
NUMA node2 CPU(s):               24-35
NUMA node3 CPU(s):               36-47
NUMA node4 CPU(s):               48-59
NUMA node5 CPU(s):               60-71
NUMA node6 CPU(s):               72-83
NUMA node7 CPU(s):               84-95
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
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 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 intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced 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 arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req 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 avx512_fp16 flush_l1d arch_capabilities

Versions of relevant libraries:
[pip3] flashinfer==0.1.4+cu121torch2.4
[pip3] numpy==1.26.4
[pip3] nvidia-cublas-cu12==12.1.3.1
[pip3] nvidia-cuda-cupti-cu12==12.1.105
[pip3] nvidia-cuda-nvrtc-cu12==12.1.105
[pip3] nvidia-cuda-runtime-cu12==12.1.105
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.0.2.54
[pip3] nvidia-curand-cu12==10.3.2.106
[pip3] nvidia-cusolver-cu12==11.4.5.107
[pip3] nvidia-cusparse-cu12==12.1.0.106
[pip3] nvidia-ml-py==12.560.30
[pip3] nvidia-nccl-cu12==2.20.5
[pip3] nvidia-nvjitlink-cu12==12.6.20
[pip3] nvidia-nvtx-cu12==12.1.105
[pip3] pyzmq==26.2.0
[pip3] torch==2.4.0
[pip3] torchvision==0.19.0
[pip3] transformers==4.44.2
[pip3] triton==3.0.0
[conda] Could not collect
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.5.5@
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    GPU1    GPU2    GPU3    NIC0    NIC1    NIC2    NIC3    NIC4    NIC5    NIC6    NIC7    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      NV18    NV18    NV18    PXB     SYS     SYS     SYS     SYS     SYS     SYS     SYS     0-11    0               N/A
GPU1    NV18     X      NV18    NV18    SYS     PIX     PIX     PXB     SYS     SYS     SYS     SYS     24-35   2               N/A
GPU2    NV18    NV18     X      NV18    SYS     SYS     SYS     SYS     SYS     SYS     SYS     PXB     72-83   6               N/A
GPU3    NV18    NV18    NV18     X      SYS     SYS     SYS     SYS     SYS     SYS     SYS     PIX     72-83   6               N/A
NIC0    PXB     SYS     SYS     SYS      X      SYS     SYS     SYS     SYS     SYS     SYS     SYS
NIC1    SYS     PIX     SYS     SYS     SYS      X      PIX     PXB     SYS     SYS     SYS     SYS
NIC2    SYS     PIX     SYS     SYS     SYS     PIX      X      PXB     SYS     SYS     SYS     SYS
NIC3    SYS     PXB     SYS     SYS     SYS     PXB     PXB      X      SYS     SYS     SYS     SYS
NIC4    SYS     SYS     SYS     SYS     SYS     SYS     SYS     SYS      X      PIX     PXB     SYS
NIC5    SYS     SYS     SYS     SYS     SYS     SYS     SYS     SYS     PIX      X      PXB     SYS
NIC6    SYS     SYS     SYS     SYS     SYS     SYS     SYS     SYS     PXB     PXB      X      SYS
NIC7    SYS     SYS     PXB     PIX     SYS     SYS     SYS     SYS     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
  NIC1: mlx5_1
  NIC2: mlx5_2
  NIC3: mlx5_3
  NIC4: mlx5_4
  NIC5: mlx5_5
  NIC6: mlx5_6
  NIC7: mlx5_7

🐛 Describe the bug

I am using a lora adapter based on the llama-65b.

I have started vllm v0.5.5 openai entrypoint with the following settings.

    - args:
      - --model
      - /data/models/llama-65b-instruct/base
      - --tensor-parallel-size
      - "4"
      - --load-format
      - "auto"
      - --max-model-len
      - "8192"
      - --enable-lora
      - --max-loras
      - "8"
      - --max-cpu-loras
      - "30"
      - --max-lora-rank
      - "64"
      - --lora-dtype
      - bfloat16
      - --uvicorn-log-level
      - warning
      - --lora-modules
      - 'lora1=/data/models/llama-65b-instruct/adapter/lora1'
      - 'lora2=/data/models/llama-65b-instruct/adapter/lora2'

adapter_config.json

{
  "alpha_pattern": {},
  "auto_mapping": null,
  "base_model_name_or_path": "/data/models/llama-65b-instruct/base",
  "bias": "none",
  "fan_in_fan_out": false,
  "inference_mode": true,
  "init_lora_weights": true,
  "layer_replication": null,
  "layers_pattern": null,
  "layers_to_transform": null,
  "loftq_config": {},
  "lora_alpha": 32,
  "lora_dropout": 0.05,
  "megatron_config": null,
  "megatron_core": "megatron.core",
  "modules_to_save": null,
  "peft_type": "LORA",
  "r": 16,
  "rank_pattern": {},
  "revision": null,
  "target_modules": [
    "q_proj",
    "k_proj",
    "v_proj",
    "o_proj"
  ],
  "task_type": "CAUSAL_LM",
  "use_dora": false,
  "use_rslora": false
}

I made 1 request with the following parameters.

    “n": 1,
    “max_tokens": 1000,
    “temperature": 0,
    “best_of": 1,
    “stream": false

1. adapter: the answer keeps changing slightly.

2. base model: vllm always guarantee the same answer when temperature is 0.

3. adapter with --fully-sharded-loras: If we add the --fully-sharded-loras option to the vllm startup, vllm always give the same answer when temperature is 0.

I got 10 answers for each condition.

What makes the difference in the answer?

About this issue

  • Original URL
  • State: open
  • Created 2 years ago
  • Comments: 24 (1 by maintainers)

Most upvoted comments

same issue vllm v0.6.4.post1, Qwen2.5-32B-Instruct-AWQ, tp=2 , temperature=0, base model greedy search got same answers lora model got different answers add --fully-sharded-loras(affect the perfermance), lora model also got different answers any idea?

facing the same issue. has anyone resolved this? how to do greedy sampling with qlora?

Nope. I haven’t solved this issue yet. However, considering precision, the LoRA operation result and the Merged model operation result will be the same only assuming that precision is infinite. Because the types and number of operations are more when using LoRA. It’s a structure in which a kind of encoder-decoder is added, and the result is added to the calculation result of the base model. In this process, minute losses will inevitably occur due to the precision of the facet, and the more layers stacked, the more differences will inevitably exist. So, if we demanded greedy sampling and the decision bondary was very narrow, so the model was in conflict over what to choose between multiple tokens, we decided that even if greedy sampling was performed, the reasoning results of LoRA and merged models may not be the same. I’d like to discuss whether my logic is correct.

Additionally, I found that this issue seems to be related to max_num_seqs. When max_num_seqs is set to 1, the greedy strategy remains consistent.

This doesn’t work for me. And setting this will cause the inference to be extremely slow (~15x - ~20x) for me.

facing the same issue. has anyone resolved this? how to do greedy sampling with qlora?

Nope. I haven’t solved this issue yet. However, considering precision, the LoRA operation result and the Merged model operation result will be the same only assuming that precision is infinite. Because the types and number of operations are more when using LoRA. It’s a structure in which a kind of encoder-decoder is added, and the result is added to the calculation result of the base model.

In this process, minute losses will inevitably occur due to the precision of the facet, and the more layers stacked, the more differences will inevitably exist.

So, if we demanded greedy sampling and the decision bondary was very narrow, so the model was in conflict over what to choose between multiple tokens, we decided that even if greedy sampling was performed, the reasoning results of LoRA and merged models may not be the same.

I’d like to discuss whether my logic is correct.

Additionally, I found that this issue seems to be related to max_num_seqs. When max_num_seqs is set to 1, the greedy strategy remains consistent.

A temporary solution I found is to merge the LoRA module into the original model, so I don’t have to use vLLM’s LoRA adapter. After doing this, the consistency was maintained. (reference https://llamafactory.readthedocs.io/zh-cn/latest/getting_started/merge_lora.html)

Of course, I know how to do that, but if there are many fine-tuned adapters for each use case, it is impossible to merge adapters and serve them separately.It will probably cost a lot of money. Theoretically, the answers of the LoRA adapter and base model should be the same, but if the premise is not established, LoRA could not be used, so I think this is an important issue.