vllm [Bug]: PaliGemma serving

Your current environment

Collecting environment information...
PyTorch version: 2.3.1+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 3.22.1
Libc version: glibc-2.35

Python version: 3.11.9 (main, Apr 19 2024, 16:48:06) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-6.1.0-21-amd64-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 A800 80GB PCIe
Nvidia driver version: 525.147.05
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:                        43 bits physical, 48 bits virtual
Byte Order:                           Little Endian
CPU(s):                               24
On-line CPU(s) list:                  0-23
Vendor ID:                            AuthenticAMD
Model name:                           AMD EPYC 7352 24-Core Processor
CPU family:                           23
Model:                                49
Thread(s) per core:                   1
Core(s) per socket:                   24
Socket(s):                            1
Stepping:                             0
Frequency boost:                      enabled
CPU max MHz:                          2300.0000
CPU min MHz:                          1500.0000
BogoMIPS:                             4600.10
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 nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sev sev_es
Virtualization:                       AMD-V
L1d cache:                            768 KiB (24 instances)
L1i cache:                            768 KiB (24 instances)
L2 cache:                             12 MiB (24 instances)
L3 cache:                             128 MiB (8 instances)
NUMA node(s):                         1
NUMA node0 CPU(s):                    0-23
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:               Mitigation; untrained return thunk; SMT disabled
Vulnerability Spec rstack overflow:   Mitigation; SMT disabled
Vulnerability Spec store bypass:      Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:             Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:             Mitigation; Retpolines; IBPB conditional; 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] lion-pytorch==0.1.4
[pip3] numpy==1.26.4
[pip3] nvidia-nccl-cu12==2.20.5
[pip3] onnx==1.16.0
[pip3] onnx-graphsurgeon==0.5.2
[pip3] onnxruntime==1.16.3
[pip3] onnxruntime-gpu==1.17.1
[pip3] open-clip-torch==2.24.0
[pip3] pytorch-lightning==2.2.4
[pip3] torch==2.3.1
[pip3] torchaudio==2.3.1
[pip3] torchdata==0.7.1
[pip3] torchmetrics==1.3.2
[pip3] torchvision==0.18.1
[pip3] transformers==4.42.4
[pip3] triton==2.3.1
[pip3] tritonclient==2.45.0
[conda] lion-pytorch              0.1.4                    pypi_0    pypi
[conda] numpy                     1.26.4          py311h24aa872_0  
[conda] numpy-base                1.26.4          py311hbfb1bba_0  
[conda] nvidia-nccl-cu12          2.20.5                   pypi_0    pypi
[conda] open-clip-torch           2.24.0                   pypi_0    pypi
[conda] pytorch-lightning         2.2.4                    pypi_0    pypi
[conda] torch                     2.3.1                    pypi_0    pypi
[conda] torchaudio                2.3.1                    pypi_0    pypi
[conda] torchdata                 0.7.1                    pypi_0    pypi
[conda] torchmetrics              1.3.2                    pypi_0    pypi
[conda] torchvision               0.18.1                   pypi_0    pypi
[conda] transformers              4.42.4                   pypi_0    pypi
[conda] triton                    2.3.1                    pypi_0    pypi
[conda] tritonclient              2.45.0                   pypi_0    pypi
ROCM Version: Could not collect
Neuron SDK Version: N/A
vLLM Version: 0.5.2
vLLM Build Flags:
CUDA Archs: Not Set; ROCm: Disabled; Neuron: Disabled
GPU Topology:
GPU0    NIC0    NIC1    CPU Affinity    NUMA Affinity
GPU0     X      SYS     SYS     0-23            N/A
NIC0    SYS      X      PIX
NIC1    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

🐛 Describe the bug

I run the following docker container

docker run --runtime nvidia --gpus all -v ~/.cache/huggingface:/root/.cache/huggingface -p 8000:8000 --env "HUGGING_FACE_HUB_TOKEN=[MY_TOKEN_HERE]"  vllm/vllm-openai --model google/paligemma-3b-mix-224

And after it is up & running, I execute the following python code

import base64

import requests
from openai import OpenAI

# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    # defaults to os.environ.get("OPENAI_API_KEY")
    api_key=openai_api_key,
    base_url=openai_api_base,
)

models = client.models.list()
model = models.data[0].id

image_url = "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"

# Use base64 encoded image in the payload
def encode_image_base64_from_url(image_url: str) -> str:
    """Encode an image retrieved from a remote url to base64 format."""

    with requests.get(image_url) as response:
        response.raise_for_status()
        return base64.b64encode(response.content).decode('utf-8')


image_base64 = encode_image_base64_from_url(image_url=image_url)
chat_completion_from_base64 = client.chat.completions.create(
    messages=[{
        "role":
        "user",
        "content": [
            {
                "type": "text",
                "text": "USER: <image>\nWhat is the content of this image?\nASSISTANT:",
            },
            {
                "type": "image_url",
                "image_url": {
                    "url": f"data:image/jpeg;base64,{image_base64}"
                },
            },
        ],
    }],
    model=model,
)

result = chat_completion_from_base64.choices[0].message.content

print(f"Chat completion output:{result}")

This results in the following ooutput:

Chat completion output:<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>

Please help.

Jul 22 '24 11:07 arseniybelkov

Can you remove the chat template and the <image> token in your prompt and simply just do

chat_completion_from_base64 = client.chat.completions.create(
    messages=[{
        "role":
        "user",
        "content": [
            {
                "type": "text",
                "text": "What is the content of this image?",
            },
            {
                "type": "image_url",
                "image_url": {
                    "url": f"data:image/jpeg;base64,{image_base64}"
                },
            },
        ],
    }],
    model=model,
)

Jul 22 '24 23:07 ywang96

Getting the following for a simple "Hi": Hi<|im_end Hi<|im_end <|im_start tả You are a helpful assistant.

For prompts with images getting only <im_start> and <im_end> tags in response.

Jul 23 '24 06:07 utkarsh995

@ywang96 I replaced the chat template with the one you suggested. I now get:

Chat completion output:<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
<|im_end>
</|im_end>
<|im_end>
</|im_end>
<|im_end>
</|im_end>
<|im_end>
</|im_end>
</|im_end>
</|im_end>
</|im_end>
</|im_end>
</|im_end>
</|im_end>
</|im_end>
</|im_end>
</|im_end>
</|im_end>
</|im_end>
</|im_end>
</|im_end>
</|im_end>

I also started using this image since the previous one is no longer available

Jul 25 '24 09:07 arseniybelkov

@arseniybelkov In fact, I don't think PaliGemma is supposed to work with OpenAI API format since it's never instruction fine-tuned and it never came with a chat-template. Can you try this model with LLM class from vLLM and see if you're still getting the same result?

Jul 25 '24 16:07 ywang96

Sorry for delay, I've been on vacation.

from vllm import LLM

from PIL import Image
import PIL
import requests
from io import BytesIO

response = requests.get("https://upload.wikimedia.org/wikipedia/commons/thumb/1/15/Tour_Saint-Jacques_au_cr%C3%A9puscule.jpg/500px-Tour_Saint-Jacques_au_cr%C3%A9puscule.jpg")
image = Image.open(BytesIO(response.content))
    
# Refer to the HuggingFace repo for the correct format to use
prompt = "What is the content of this image?"

# Load the image using PIL.Image
llm = LLM(model="google/paligemma-3b-mix-224")
# Single prompt inference
outputs = llm.generate({
    "prompt": prompt,
    "multi_modal_data": {"image": image},
})

for o in outputs:
    generated_text = o.outputs[0].text
    print(generated_text)

This code results in this output

Processed prompts: 100%|██████████████████████████████████████████████████████| 1/1 [00:00<00:00, 22.63it/s, est. speed input: 6028.09 toks/s, output: 67.96 toks/s]
 GHFW

Jul 30 '24 09:07 arseniybelkov

@arseniybelkov I don't think the model is supposed to work that way - could you try caption es as the prompt? (This is the same example given in the documentation from huggingface, which is # Instruct the model to create a caption in Spanish)

You can refer to the example here

Aug 05 '24 06:08 ywang96

Yes, you're right, I didn't pay attention to it. I changed the prompt to "caption en" and now it gives me this

In this image there are trees, at Iliad intersection个人, at the top there

Aug 09 '24 12:08 arseniybelkov

@ywang96 ping just in case

Aug 13 '24 13:08 arseniybelkov

Yes, you're right, I didn't pay attention to it. I changed the prompt to "caption en" and now it gives me this
In this image there are trees, at Iliad intersection个人, at the top there

It does seem that the generation does start correctly...which precision is this?

Aug 14 '24 04:08 ywang96

I've just set dtype=torch.float32 and got the same result

Aug 14 '24 12:08 arseniybelkov

@arseniybelkov Is it possible if you can send me the image and the prompt so I can repro and compare to transformers (especially on input embeddings)?

Aug 14 '24 17:08 ywang96

@ywang96 Here is the code and image

from vllm import LLM

from PIL import Image
import PIL
import requests
from io import BytesIO
from transformers import AutoTokenizer
import torch


MODEL_ID = "google/paligemma-3b-mix-224"

image = Image.open("Notre-Dame_de_Paris,_4_October_2017.jpg")

# Refer to the HuggingFace repo for the correct format to use
prompt = "caption en"

# Load the image using PIL.Image
llm = LLM(model=MODEL_ID, tokenizer=MODEL_ID, dtype=torch.float32)
# Single prompt inference
outputs = llm.generate({
    "prompt": prompt,
    "multi_modal_data": {"image": image},
})

for o in outputs:
    generated_text = o.outputs[0].text
    print(generated_text)

Aug 15 '24 10:08 arseniybelkov

Hey @arseniybelkov I just want to let you know this is still on my list of TODO but I simply had limited bandwidth with other priorities. Sorry for the inconvenience!

Aug 22 '24 19:08 ywang96

This issue has been automatically marked as stale because it has not had any activity within 90 days. It will be automatically closed if no further activity occurs within 30 days. Leave a comment if you feel this issue should remain open. Thank you!

Nov 22 '24 02:11 github-actions[bot]

This issue has been automatically closed due to inactivity. Please feel free to reopen if you feel it is still relevant. Thank you!

Dec 22 '24 02:12 github-actions[bot]