[Model] Support Mamba

[tlrmchlsmth]

trafficstars

This is closely based on vLLM's Jamba implementation. It also has several changes and fixes to deal with the fact that there is no KV cache at all.

Changes in this PR

• Added the Mamba model definition and integration tests.
• Factored the Mamba cache management used by both Mamba and Jamba into a mamba_cache.py
• Added a new "placeholder" attention backend with many noop methods, as Mamba's state needs to be managed differently. AFAICT, this is an expedient way to get Mamba working without a larger refactor. We didn't need to do this for Jamba because Jamba does have some attention layers and does have an attention implementation.
• Various other changes to wrestle with the fact that Mamba doesn't have attention.

Support for Mamba2, Codestral-Mamba, and FalconMamba will come in subsequent PRs.

---

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!

[github-actions[bot]]

👋 Hi! Thank you for contributing to the vLLM project. Just a reminder: PRs would not trigger full CI run by default. Instead, it would only trigger fastcheck CI to run, which consists only a small and essential subset of tests to quickly catch errors with the flexibility to run extra individual tests on top (you can do this by unblocking test steps in the Buildkite run).

Full CI run is still required to merge this PR so once the PR is ready to go, please make sure to run it. If you need all test signals in between PR commits, you can trigger full CI as well.

To run full CI, you can do one of these:

• Comment /ready on the PR
• Add ready label to the PR
• Enable auto-merge.

🚀

[zhiyuan1i]

Thank you very much for your work, I want to implement the vllm feature for RWKV, your implementation gives me a great reference. Can I work with you to bring RWKV, an RNN model? I have a native pytorch implementation here, but it still needs to be modified, and I don't know how to adapt to so many VLLM backends (CUDA, XPU, TPU, HIP, etc.)

[tlrmchlsmth]

Hey @uniartisan, sure I can work with you to implement support for RWKV. Perhaps you could you start a draft PR, and we can talk there. Could you link your native pytorch implementation?

I think it would be a good idea to start with CUDA support -- for Mamba, for instance, we depend on kernels that require CUDA, and you may run into that as well.

[zhiyuan1i]

Hey @uniartisan, sure I can work with you to implement support for RWKV. Perhaps you could you start a draft PR, and we can talk there. Could you link your native pytorch implementation?

I think it would be a good idea to start with CUDA support -- for Mamba, for instance, we depend on kernels that require CUDA, and you may run into that as well.

I have submitted an implementation of the model code (without any adaptation for vllm; I may come back to do this in the coming week). I will create a new draft and mention you. Thank you! My draft is available here: https://github.com/vllm-project/vllm/pull/6749 It's a native PyTorch implementation. I will first try to adapt it for vllm, and then port the CUDA kernel and Triton kernel.

[anguswangrv]

This is closely based on vLLM's Jamba implementation. It also has several changes and fixes to deal with the fact that there is no KV cache at all. This PR adds a placeholder attention backend, and adapts and renames EmbeddingModelBlockManager to handle cases like Mamba as well. TODO before landing:

• [x] Fix issues with enforce_eager=True (this issue simply vanished after merging)
• [ ] Factor out common code between Jamba and Mamba
• [x] Added a unit test.

I will also try to get Codestral Mamba working as well. The transformers-compatible Mamba models seem to be working. However, the mamba2 models unfortunately are not transformers-compatible and don't work out of the box. Codestral https://huggingface.co/mistralai/mamba-codestral-7B-v0.1 has the same issue.

See also #6479

PR Checklist (Click to Expand)

Hi @tlrmchlsmth, thanks for your work! Does the current state of this PR support Codestral Mamba on vLLM yet? If so, how do you run it?

[learning-chip]

However, the mamba2 models unfortunately are not transformers-compatible and don't work out of the box.

Huggingface transformers 4.40.0 supports Mamba-2 (codestral mamba) https://github.com/huggingface/transformers/pull/32080

[tlrmchlsmth]

@anguswangrv, Unfortunately, the Mamba-Codestral released by Mistral is not transformers-compatible. You can see here https://huggingface.co/mistralai/Mamba-Codestral-7B-v0.1/tree/main, there is no config.json file and so won't work out of the box in vLLM.

@learning-chip thanks for the pointer. Given that https://github.com/huggingface/transformers/pull/32080 added modeling_mamba2.py, I'll likely need do some additional work to get mamba2 working in vLLM. If anyone knows of a transformers-compatible mamba2 on huggingface, that would help me to get it working.

[molbap]

Hi @tlrmchlsmth , codestral is indeed compatible with HF transformers! I uploaded the converted files, there is a PR open on refs/pr/9, you can use it as such

from transformers import Mamba2ForCausalLM, AutoTokenizer
import torch
model_id = 'mistralai/Mamba-Codestral-7B-v0.1'
tokenizer = AutoTokenizer.from_pretrained(model_id, revision='refs/pr/9', from_slow=True, legacy=False)
model = Mamba2ForCausalLM.from_pretrained(model_id, revision='refs/pr/9', torch_dtype=torch.bfloat16)
input_ids = tokenizer("Hey how are you doing?", return_tensors= "pt")["input_ids"]

out = model.generate(input_ids, max_new_tokens=10)
print(tokenizer.batch_decode(out))

So you just need to check out the correct revision and it should work :) edit: added precision parameter to code snippet

[tlrmchlsmth]

@molbap Very nice! I will get that working in this PR :)

[tlrmchlsmth]

Going to add support for Mamba2 and Mamba-Codestral in a subsequent PR partially due to interactions with https://github.com/vllm-project/vllm/pull/7651, and partially to minimize the changes in this PR. I do have it working in another branch, so can follow pretty quickly.

[DarkLight1337]

A quick heads-up that the new locations of the model tests have been adjusted in #7820, so please merge from main.

[DarkLight1337]

Remember to add this to the Supported Models page!