-
-
Notifications
You must be signed in to change notification settings - Fork 4.1k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[Model] Make llama3.2 support multiple and interleaved images #9095
base: main
Are you sure you want to change the base?
Conversation
👋 Hi! Thank you for contributing to the vLLM project. Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can do one of these:
🚀 |
c57ce74
to
64acbc5
Compare
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can you add some unit tests on multiple image & interleaved image?
vllm/model_executor/models/mllama.py
Outdated
attn_metadata, | ||
attn_type=AttentionType.ENCODER_DECODER) | ||
if attention_mask is not None: | ||
if len(kv_cache.shape) == 3: |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Why do we need this check?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
it's for skipping writing kv-cache for the initial profiling run
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@ywang96 Is there a better way to know whether we are in the profile run?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Not really from within the forward pass, so this check is okay imo. Let's add a NOTE here to indicate the purpose of this check
Based on personal discussion with @xiangxu-google , we still need the following steps:
|
64acbc5
to
e9af2a8
Compare
e9af2a8
to
6aa5a8d
Compare
I talked to @heheda12345 offline about the following steps for this model:
from transformers import AutoTokenizer
model_id = "meta-llama/Llama-3.2-11B-Vision-Instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id)
messages = [
{"role": "user", "content": [
{"type": "image"},
{"type": "text", "text": "Describe this image"}
]},
{"role": "assistant", "content": [
{"type": "text", "text": "This image is xxx"}
]},
{"role": "user", "content": [
{"type": "image"},
{"type": "text", "text": "How about this image"}
]}
]
input_text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
print(input_text) Output:
|
Curious why do we need the logic for each request? This PR has implemented computing masks for a batch of requests. |
You're right, and it's really a matter of where we put that mask computation logic. IMO it's cleaner if we decouple the mask computation for each individual requests and what's needed for the batch in the forward pass, since this is more or less still a CPU operation that we should put in I'm also okay with leaving this logic as a whole for the batch inside |
FYI I tried using this PR to evaluate Llama 3.2 Vision 11B on MMMU and ran into a CUDA error during attention
|
What is your eval environment, like torch version, vllm version, GPU version, CUDA version. I can't reproduce the issue on H100 CUDA 12.4 with vllm installed from source. q = q.transpose(0, 1).view(self.num_key_value_groups,
self.num_local_key_value_heads, q_len,
self.head_dim)
k = k.transpose(0, 1).expand(self.num_key_value_groups,
self.num_local_key_value_heads, kv_len,
self.head_dim)
v = v.transpose(0, 1).expand(self.num_key_value_groups,
self.num_local_key_value_heads, kv_len,
self.head_dim)
attention_mask = attention_mask.view(1, 1, q_len, kv_len)
output = F.scaled_dot_product_attention(q,
k,
v,
attn_mask=attention_mask) |
Thanks for the suggestion! I need to make sure the correctness of this implementation first for a controllable landing. We can refactor it in a separate PR to move the CPU logic to |
FILL IN THE PR DESCRIPTION HERE
FIX #xxxx (link existing issues this PR will resolve)
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:
format.sh
to format your code.docs/source/
if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.Adding or changing kernels
Each custom kernel needs a schema and one or more implementations to be registered with PyTorch.
Tensors
require meta-functions. Meta-functions should be implemented and registered in python so that dynamic dims can be handled automatically. See above documents for a description of meta-functions.torch.libary.opcheck()
to test the function registration and meta-function for any registered ops. Seetests/kernels
for examples.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:
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.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!