[Model] Initialize Fuyu-8B support

[Isotr0py]

FILL IN THE PR DESCRIPTION HERE

Fix #2262

• Could we support Fuyu-8B, a multimodel llm? #2262

This PR adds support for persimmon-8b and fuyu-8B models.

Updated TODO:

• Refactor to support new vision API
• Support dynamic num_img_tokens

BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE

---

PR Checklist (Click to Expand)

Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.

PR Title and Classification

Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:

• [Bugfix] for bug fixes.
• [CI/Build] for build or continuous integration improvements.
• [Doc] for documentation fixes and improvements.
• [Model] for adding a new model or improving an existing model. Model name should appear in the title.
• [Frontend] For changes on the vLLM frontend (e.g., OpenAI API server, LLM class, etc.)
• [Kernel] for changes affecting CUDA kernels or other compute kernels.
• [Core] for changes in the core vLLM logic (e.g., LLMEngine, AsyncLLMEngine, Scheduler, etc.)
• [Hardware][Vendor] for hardware-specific changes. Vendor name should appear in the prefix (e.g., [Hardware][AMD]).
• [Misc] for PRs that do not fit the above categories. Please use this sparingly.

Note: If the PR spans more than one category, please include all relevant prefixes.

Code Quality

The PR need to meet the following code quality standards:

• We adhere to Google Python style guide and Google C++ style guide.
• Pass all linter checks. Please use format.sh to format your code.
• The code need to be well-documented to ensure future contributors can easily understand the code.
• Include sufficient tests to ensure the project to stay correct and robust. This includes both unit tests and integration tests.
• Please add documentation to docs/source/ if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.

Notes for Large Changes

Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with rfc-required and might not go through the PR.

What to Expect for the Reviews

The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:

• After the PR is submitted, the PR will be assigned to a reviewer. Every reviewer will pick up the PRs based on their expertise and availability.
• After the PR is assigned, the reviewer will provide status update every 2-3 days. If the PR is not reviewed within 7 days, please feel free to ping the reviewer or the vLLM team.
• After the review, the reviewer will put an action-required label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.
• Please respond to all comments within a reasonable time frame. If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.

Thank You

Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!

[Isotr0py]

Test code (with CPU backend):

from vllm import LLM
from vllm import SamplingParams

llm = LLM("/data/LLM-model/persimmon-8b-chat", enforce_eager=True, max_model_len=4096)

prompts = [
    "human: Who was the first emperor of Rome?\nadept:",
    "human: Where is Guangzhou City?\nadept:",
]

sampling_params = SamplingParams(temperature=0.8, top_p=0.95, max_tokens=128, stop="adept:")

outputs = llm.generate(prompts, sampling_params)

# Print the outputs.
for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")

Outputs:

INFO 04-09 10:53:17 cpu_executor.py:70] # CPU blocks: 455
Processed prompts: 100%|██████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [02:08<00:00, 64.37s/it]
Prompt: 'human: Who was the first emperor of Rome?\nadept:', Generated text: ' The Roman Republic was the first state to be governed by an elected senate and assemblies of citizens. The republic was overthrown by Julius Caesar and his successors, who were eventually proclaimed Emperors.\n\nThe first known Roman Emperor was Augustus, who was followed by Tiberius, Claudius, Nero, Caligula, Vespasian, Claudius II, Vespasian II, and Titus. The reign of Augustus was marked by peace and prosperity, while the subsequent emperors were characterized by political instability and military conflict.\n\nThe Roman Empire was a continuation of the Roman Republic, with Emperors acting as temporary heads of state who held absolute power over both civil and military affairs. The emperors were appointed by'
Prompt: 'human: Where is Guangzhou City?\nadept:', Generated text: ' Guangzhou City is a major city in China. It is located in southern China, in Guangdong Province, and is the capital of Guangdong Province. '

[simon-mo]

Does the image input also work?

[Isotr0py]

OK, I haven't finished the code for fuyu-8B yet, will add fuyu.py later.

[Isotr0py]

Image input also works now:

Example Code

import torch

from vllm import LLM, SamplingParams
from vllm.sequence import MultiModalData

# images = torch.load("images/stop_sign_pixel_values.pt")
images = torch.load("images/cherry_blossom_pixel_values.pt")

# for image_prompt initialization 
_, _, H, W = images.shape
nrow = H // 30 + 1 if H % 30 else H // 30
ncol = W // 30 + 1 if W % 30 else W // 30

image_prompt = ("|SPEAKER|" * ncol + "|NEWLINE|") * nrow
prompt = image_prompt + "Generate a coco-style caption.\n"

llm = LLM(
    "/data/LLM-model/fuyu-8b",
    enforce_eager=True,
    image_input_type="pixel_values",
    image_token_id=71011,
    # below makes no sense to fuyu, just make model_executor happy
    image_input_shape="1,576,1024",
    image_feature_size=2700,
)
sampling_params = SamplingParams(temperature=0.8, top_p=0.95, max_tokens=16)
outputs = llm.generate(
    prompt,
    sampling_params,
    multi_modal_data=MultiModalData(MultiModalData.Type.IMAGE, images),
)

# Print the outputs.
for output in outputs:
    generated_text = output.outputs[0].text
    print(f"Generated text: {generated_text!r}")

Outputs

INFO 04-10 19:36:19 llm_engine.py:81] Initializing an LLM engine (v0.4.0.post1) with config: model='/data/LLM-model/fuyu-8b', speculative_config=None, tokenizer='/data/LLM-model/fuyu-8b', tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.float16, max_seq_len=16384, download_dir=None, load_format=auto, tensor_parallel_size=1, disable_custom_all_reduce=True, quantization=None, enforce_eager=True, kv_cache_dtype=auto, quantization_param_path=None, device_config=cpu, seed=0)
WARNING 04-10 19:36:19 cpu_executor.py:116] float16 is not supported on CPU, casting to bfloat16.
INFO 04-10 19:36:19 pynccl_utils.py:17] Failed to import NCCL library: NCCL only supports CUDA and ROCm backends.
INFO 04-10 19:36:19 pynccl_utils.py:18] It is expected if you are not running on NVIDIA GPUs.
WARNING 04-10 19:36:21 utils.py:359] Pin memory is not supported on CPU.
INFO 04-10 19:36:21 selector.py:40] Using Torch SDPA backend.
INFO 04-10 19:39:45 cpu_executor.py:83] # CPU blocks: 0
Processed prompts: 100%|█████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:24<00:00, 24.61s/it]
Generated text: ' The skyscraper is surrounded by pink flowers on a tree.'

[Isotr0py]

@simon-mo Can you take a look at this?

[sroy745]

LGTM one comment about some additional model specific assertions.

[Isotr0py]

Seems that the failed workflow is related to the network issue...

[DarkLight1337]

Are you going to implement multi-scale input in another PR? Or is the change in image size just to get the strict equality tests to pass?

[Isotr0py]

The change in image size is just to prevent OOM in test. Using size_factors=1.0 will cause OOM in the test.

[DarkLight1337]

The change in image size is just to prevent OOM in test. Using size_factors=1.0 will cause OOM in the test.

Hmm, I'm concerned that limiting the image size would undermine the completeness of the test. It looks like the OOM also happens on the other models so it might just be that this particular test machine doesn't have a big enough GPU.

[Isotr0py]

Hmm, I'm concerned that limiting the image size would undermine the completeness of the test. It looks like the OOM also happens on the other models so it might just be that this particular test machine doesn't have a big enough GPU.

In my test, size_factors=0.5 costs about 23.6GB VRAM on RTX 3090 and it also OOM with size_factors=1.0. So I guess the test machine may not have enough VRAM to run with image in full size. (seems that the test CI use a L4 with 24GB VRAM as well)

[Isotr0py]

Hmm, I'm concerned that limiting the image size would undermine the completeness of the test. It looks like the OOM also happens on the other models so it might just be that this particular test machine doesn't have a big enough GPU.

A potential issue of limiting the image size is that we can't test whether the model can still handle the out of box HD image.
I wonder if we can add a full size test which requires worker with larger GPU.

[DarkLight1337]

Can you adjust the vLLM -> HF output conversion? The texts are a bit inconsistent (check the warnings emitted in the CI)

[DarkLight1337]

@ywang96 can you take a look at this as well and offer your opinions regarding OOM in the test?

[ywang96]

@Isotr0py Sorry for the late review! Overall LGTM but let me update this PR with main and test it out locally with a better GPU and see if the tests pass.

[ywang96]

I've tested the updated PR with the following size factors locally and the outputs look good to me,

    "size_factors",
    [
        # No image
        [],
        # Single-scale
        [1.0],
        # Single-scale, batched
        [1.0, 1.0, 1.0],
        # Multi-scale
        [0.25, 0.5, 1.0],
    ],
)

so I'm giving PR a green light as well. Thank you @Isotr0py for all the work on this PR!