Skip to content

[Bugfix] fix pp for llama4#16746

Merged
DarkLight1337 merged 2 commits intovllm-project:mainfrom
luccafong:llama4-pp-fix
Apr 18, 2025
Merged

[Bugfix] fix pp for llama4#16746
DarkLight1337 merged 2 commits intovllm-project:mainfrom
luccafong:llama4-pp-fix

Conversation

@luccafong
Copy link
Collaborator

@luccafong luccafong commented Apr 17, 2025 •
edited by github-actions bot
Loading

This PR fixes PP for llama4 (#16385)

Loading language model with architecture to support PP verification and correct the prefix to separate model loading weights for language model and rest of the models so both PP=0 and PP=1 can work.

VLLM_DISABLE_COMPILE_CACHE=1 python -m vllm.entrypoints.openai.api_server --model $LLAMA_DIR --served-model-name maverick  --max-model-len 20000 --tensor-parallel 4 --pipeline-parallel-size 2 --gpu-memory-utilization 0.95  --port 8000 

curl http://localhost:8000/v1/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "maverick",
        "prompt": "San Francisco is a",
        "max_tokens": 7,
        "temperature": 0
    }'

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.930|±  |0.0181|
|     |       |strict-match    |     5|exact_match|↑  |0.935|±  |0.0175|

@github-actions
Copy link

github-actions bot commented Apr 17, 2025

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

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 either: Add ready label to the PR or enable auto-merge.

🚀

@luccafong luccafong marked this pull request as ready for review April 17, 2025 01:16
@luccafong
fix pp for llama4
fdc4236 
Signed-off-by: Lu Fang <fanglu@fb.com>
@luccafong luccafong requested a review from houseroad April 17, 2025 03:56
@luccafong luccafong mentioned this pull request Apr 17, 2025
Closed
1 task
@luccafong luccafong requested a review from ywang96 April 17, 2025 04:03
houseroad
houseroad reviewed Apr 17, 2025
View reviewed changes
vllm/model_executor/models/mllama4.py
self.language_model = _initialize_model(
vllm_config=vllm_config.with_hf_config(config.text_config),
vllm_config=vllm_config.with_hf_config(config.text_config,
["LlamaForCausalLM"]),
Copy link
Collaborator

@houseroad houseroad Apr 17, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should we use Llama4ForCausalLM?

Copy link
Collaborator Author

@luccafong luccafong Apr 17, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Llama4ForCasualLm is not registered architecture, we should avoid using that which requires adding a lot of hacks as in the initial PR

Copy link
Member

@ywang96 ywang96 Apr 17, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Probably a dumb question, but since _initialize_model is already pointing to model_class=Llama4ForCausalLM, why do we need to override the architectures here to LlamaForCausalLM?

vllm/vllm/model_executor/model_loader/loader.py

Lines 114 to 123 in 8cac35b

def _initialize_model(
vllm_config: VllmConfig,
*,
prefix: str = "",
model_class: Optional[type[nn.Module]] = None,
) -> nn.Module:
"""Initialize a model with the given configurations."""
model_config = vllm_config.model_config
if model_class is None:
model_class, _ = get_model_architecture(model_config)

Copy link
Collaborator Author

@luccafong luccafong Apr 17, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

during __post__init__ of the hf config when we call replace inside with_hf_config function

vllm/vllm/config.py

Line 3781 in a018e55

return replace(self, model_config=model_config)
, it will check if the architecture is supporting PP or not via

vllm/vllm/config.py

Line 3790 in a018e55

self.model_config.verify_with_parallel_config(self.parallel_config)
, where None architecture will raise an issue as in #16385 in below code

vllm/vllm/model_executor/models/registry.py

Lines 441 to 442 in a018e55

normalized_arch = list(
filter(lambda model: model in self.models, architectures))

houseroad
houseroad reviewed Apr 17, 2025
View reviewed changes
vllm/model_executor/models/mllama4.py
# using llama4's load_weights routine.
language_model_weights, other_weights = self.separate_weights(
weights, prefix="language_model.model.")
weights, prefix="language_model.")
Copy link
Collaborator

@houseroad houseroad Apr 17, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Wondering why this issue was not triggered before?

Copy link
Collaborator Author

@luccafong luccafong Apr 17, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

language_model.lm_head can also be loaded by the parent model weight loader if not PP enabled, but for PP, as we split the weights into 2 parts, the lm_head is missing in PP=0 so it will raise an issue weight not found. while in llama4.py model loading we have logic handling is_pp_missing_parameter to avoid the exception

houseroad
houseroad approved these changes Apr 17, 2025
View reviewed changes
Copy link
Collaborator

@houseroad houseroad left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for the fix.

@houseroad houseroad added the ready ONLY add when PR is ready to merge/full CI is needed label Apr 17, 2025
@DarkLight1337 DarkLight1337 merged commit e31045f into vllm-project:main Apr 18, 2025
62 checks passed
yangw-dev pushed a commit to yangw-dev/vllm that referenced this pull request Apr 21, 2025
@luccafong @yangw-dev
[Bugfix] fix pp for llama4 (vllm-project#16746)
9f8a7e5 
Signed-off-by: Lu Fang <fanglu@fb.com>
Signed-off-by: Yang Wang <elainewy@meta.com>
jikunshang pushed a commit to jikunshang/vllm that referenced this pull request Apr 29, 2025
@luccafong
[Bugfix] fix pp for llama4 (vllm-project#16746)
309ef50 
Signed-off-by: Lu Fang <fanglu@fb.com>
lk-chen pushed a commit to lk-chen/vllm that referenced this pull request Apr 29, 2025
@luccafong @lk-chen
[Bugfix] fix pp for llama4 (vllm-project#16746)
b30d0e2 
Signed-off-by: Lu Fang <fanglu@fb.com>
adobrzyn pushed a commit to HabanaAI/vllm-fork that referenced this pull request Apr 30, 2025
@luccafong @adobrzyn
[Bugfix] fix pp for llama4 (vllm-project#16746)
4984e14 
Signed-off-by: Lu Fang <fanglu@fb.com>
Signed-off-by: Agata Dobrzyniewicz <adobrzyniewicz@habana.ai>
RichardoMrMu pushed a commit to RichardoMrMu/vllm that referenced this pull request May 12, 2025
@luccafong
[Bugfix] fix pp for llama4 (vllm-project#16746)
4e201ae 
Signed-off-by: Lu Fang <fanglu@fb.com>
Signed-off-by: Mu Huai <tianbowen.tbw@antgroup.com>
@ckhordiasma ckhordiasma mentioned this pull request May 14, 2025
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@houseroad houseroad houseroad approved these changes

@ywang96 ywang96 Awaiting requested review from ywang96

Assignees

No one assigned

Labels

ready ONLY add when PR is ready to merge/full CI is needed

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@luccafong @houseroad @ywang96 @DarkLight1337