[Core] Enable Memory Tiering for vLLM

[PanJason]

This is the following PR of #7697. This PR introduces 3 major functionalities:

1. Context Caching with TTL support is also a huge drive for memory demand.
2. Blockwise swapping for DRAM and Disk (Partially)
3. Layered transfer between DRAM and HBM.
4. Initial Disk Support

Some detailed explanations:

1. Context caching: a. Usage: We add CachingParams to specify the TTL of the context caching. Within TTL we guarantee that the Context Caching will remain in the instance (won't be discarded) Example usage of offline inference:

caching_params = CachingParams(ttl=args.ttl)
cache_output = llm.caching(LONG_PROMPT, caching_params=caching_params)

Example usage of online serving (a new endpoint is added: v1/caching):

 url += "/v1/caching"
 payload["ttl"] = ttl

and send the payload to the URL

b. Implementation: The scheduler.schedule() checks and frees the expired context caching LLMEngine._process_model_outputs() moves the context caching requests to SequenceStatues.FIXED, which won't be freed within TTL

2. Blockwise swapping for DRAM and Disk (partially) Usage: Add --enable-memory-tiering flag

Implementation: Prefix matching is changed in 2 places: 1. can_allocate() now considers the prefix matching to calculate the necessary new blocks 2. In prefilling, we check whether the block resides in DRAM or Disk if so we fetch them in

Block allocation is changed for prefilling and decoding. When we allocate a new block that was computed, we check whether there is free space in DRAM and disk. We swap the block out if there is enough space. Note that this can also have a cascading effect (e.g. GPU -> CPU, CPU -> Disk) which we also handle

rmap is introduced to track which seq is mapped to the block since the context caching blocks have to exist during the TTL. The context caching only blocks are swappable (because if they are not used they will occupy the whole HBM) so we need to find the sequences that mapped to a certain block in eviction.

3. Layered transfer between HBM and DRAM: Usage: Add --enable-layered-transfer flag

Implementation: We create another CUDA stream for transferring. In attention calculation, the calculation waits for this dedicated CUDA stream and invokes the data transfer for the next layer in this dedicated CUDA stream. Then, it starts calculating the current layer.

4. Disk support (Temtative) Usage: Add --enable-disk-swap flag and pass the config file by --disk-swap-config swap.cfg

Implementation Right now disk still uses block abstract and the same granularity. The swap manager is separated into the stateful part (SwapDeviceManager) and stateless part (SwapDeviceClient) because the scheduler and the works use different processes, and it is hard to share states between them (we can do it. It Just takes more time). The stateful part manages the block allocation which the stateless part is responsible for transmission (Like the separation between CacheEngine and BlockManager). Right now each block and each layer will create a file. I will optimize it later.

Supporting status:

1. Only BlockManagerV1 is supported. BlockManagerV2 is WIP.
2. Disk does not support layered transfer now as I am still exploring the interface
3. Disk only supports TP=n case as PP=n requires sharing between multiple schedulers.

---

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.

Adding or changing kernels

Each custom kernel needs a schema and one or more implementations to be registered with PyTorch.

• Make sure custom ops are registered following PyTorch guidelines: Custom C++ and CUDA Operators and The Custom Operators Manual
• Custom operations that return 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.
• Use torch.libary.opcheck() to test the function registration and meta-function for any registered ops. See tests/kernels for examples.
• When changing the C++ signature of an existing op, the schema must be updated to reflect the changes.
• If a new custom type is needed, see the following document: Custom Class Support in PT2.

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!