Improved prefix caching for paged attention

[EricLBuehler]

Summary

• fix LogicalTokenBlock::pop_token so block size stays constant
• re-enable paged prefix caching now that it works

Testing

• cargo test -p mistralrs-core --no-run (fails: extern location for darling_core does not exist)

---

https://chatgpt.com/codex/tasks/task_e_6841cd341b30832290902dc473c3c3f4

Summary by CodeRabbit

• Bug Fixes

  • Improved internal token management to ensure more reliable token removal and memory handling.

• New Features

  • Enabled functional caching for sequences backed by the block engine, ensuring proper cache creation and reference count management.
  • Enhanced block caching and matching logic for better handling of token blocks and offsets in paged-attention sequences.

[coderabbitai[bot]]

[!IMPORTANT]

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

"""

Walkthrough

Internal logic for the pop_token method in the LogicalTokenBlock struct was changed to clear tokens without shrinking the tokens vector in two modules. Additionally, the add_sequence method in PrefixCacheManagerV2 was updated to fully implement block-engine-backed caching, replacing previously commented-out code with active logic for cache creation, reference counting, and improved handling of logical token blocks and token offsets.

Changes

File(s)	Change Summary
mistralrs-core/src/dummy_paged_attention/block_engine.rs, mistralrs-core/src/paged_attention/block_engine.rs	Modified LogicalTokenBlock::pop_token to clear the last token by setting it to zero instead of removing it from the vector.
mistralrs-core/src/prefix_cacher.rs	Activated and refined block caching logic in PrefixCacheManagerV2::add_sequence; enhanced search_for_matching_cache to manage logical blocks and token offsets for paged-attention caches.

Sequence Diagram(s)

sequenceDiagram
    participant User
    participant PrefixCacheManagerV2
    participant BlockEngine
    participant BlockCaches

    User->>PrefixCacheManagerV2: add_sequence(sequence)
    PrefixCacheManagerV2->>BlockEngine: get_mut_arcmutex()
    PrefixCacheManagerV2->>BlockEngine: access block table for sequence_id
    BlockEngine-->>PrefixCacheManagerV2: block_table or panic if not found
    PrefixCacheManagerV2->>PrefixCacheManagerV2: hash logical token blocks
    PrefixCacheManagerV2->>BlockCaches: check if hash exists
    alt Not cached
        PrefixCacheManagerV2->>BlockEngine: increment ref count for each block
        PrefixCacheManagerV2->>BlockCaches: insert new BlockCacheElement
    end

Possibly related PRs

• EricLBuehler/mistral.rs#1359: Refactors PrefixCacheManagerV2 cache data structures and search approach, affecting the same module and cache management logic as this PR.

Poem

A hop and a skip through the cache we go,
Clearing tokens, not shrinking the row.
Blocks are counted, caches are made,
No more commented code in the shade!
With every pop, the logic stays neat—
A rabbit’s work is always discrete.
🐇✨ """

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

🪧 Tips

Chat

There are 3 ways to chat with CodeRabbit:

• Review comments: Directly reply to a review comment made by CodeRabbit. Example:
  • I pushed a fix in commit <commit_id>, please review it.
  • Explain this complex logic.
  • Open a follow-up GitHub issue for this discussion.
• Files and specific lines of code (under the "Files changed" tab): Tag @coderabbitai in a new review comment at the desired location with your query. Examples:
  • @coderabbitai explain this code block.
  • @coderabbitai modularize this function.
• PR comments: Tag @coderabbitai in a new PR comment to ask questions about the PR branch. For the best results, please provide a very specific query, as very limited context is provided in this mode. Examples:
  • @coderabbitai gather interesting stats about this repository and render them as a table. Additionally, render a pie chart showing the language distribution in the codebase.
  • @coderabbitai read src/utils.ts and explain its main purpose.
  • @coderabbitai read the files in the src/scheduler package and generate a class diagram using mermaid and a README in the markdown format.
  • @coderabbitai help me debug CodeRabbit configuration file.

Support

Need help? Create a ticket on our support page for assistance with any issues or questions.

Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments.

CodeRabbit Commands (Invoked using PR comments)

• @coderabbitai pause to pause the reviews on a PR.
• @coderabbitai resume to resume the paused reviews.
• @coderabbitai review to trigger an incremental review. This is useful when automatic reviews are disabled for the repository.
• @coderabbitai full review to do a full review from scratch and review all the files again.
• @coderabbitai summary to regenerate the summary of the PR.
• @coderabbitai generate docstrings to generate docstrings for this PR.
• @coderabbitai generate sequence diagram to generate a sequence diagram of the changes in this PR.
• @coderabbitai resolve resolve all the CodeRabbit review comments.
• @coderabbitai configuration to show the current CodeRabbit configuration for the repository.
• @coderabbitai help to get help.

Other keywords and placeholders

• Add @coderabbitai ignore anywhere in the PR description to prevent this PR from being reviewed.
• Add @coderabbitai summary to generate the high-level summary at a specific location in the PR description.
• Add @coderabbitai anywhere in the PR title to generate the title automatically.

CodeRabbit Configuration File (.coderabbit.yaml)

• You can programmatically configure CodeRabbit by adding a .coderabbit.yaml file to the root of your repository.
• Please see the configuration documentation for more information.
• If your editor has YAML language server enabled, you can add the path at the top of this file to enable auto-completion and validation: # yaml-language-server: $schema=https://coderabbit.ai/integrations/schema.v2.json

Documentation and Community

• Visit our Documentation for detailed information on how to use CodeRabbit.
• Join our Discord Community to get help, request features, and share feedback.
• Follow us on X/Twitter for updates and announcements.

[github-actions[bot]]

Code Metrics Report

===============================================================================
 Language            Files        Lines         Code     Comments       Blanks
===============================================================================
 C Header                3           62           53            0            9
 CSS                     1          473          408           14           51
 Dockerfile              1           42           23           10            9
 HTML                    1           73           61            4            8
 JavaScript              7         1248          936          174          138
 JSON                   14          123          122            0            1
 Makefile                1            6            5            0            1
 Python                 87         4097         3457          161          479
 Shell                   1           63           26           18           19
 Plain Text              3         3723            0         2413         1310
 TOML                   21          695          634           10           51
 YAML                    2           21           19            2            0
-------------------------------------------------------------------------------
 Jupyter Notebooks       3            0            0            0            0
 |- Markdown             2           77           32           31           14
 |- Python               2          205          178            1           26
 (Total)                            282          210           32           40
-------------------------------------------------------------------------------
 Markdown               60         5211            0         3984         1227
 |- BASH                11          123          117            2            4
 |- JSON                 2           42           42            0            0
 |- Python               7          121          109            0           12
 |- Rust                22          757          634            1          122
 |- TOML                 2           75           63            0           12
 (Total)                           6329          965         3987         1377
-------------------------------------------------------------------------------
 Rust                  376       132481       117880         2913        11688
 |- Markdown           175         3002           29         2662          311
 (Total)                         135483       117909         5575        11999
===============================================================================
 Total                 581       148318       123624         9703        14991
===============================================================================

[sempervictus]

@EricLBuehler - i was hoping this might be the silver bullet to the OOMs i've been seeing but unfortunately i was able to reproduce without paged attention at all yesterday. That said, having our own MMU approximation with ref counts and awareness of utilization should serve to prevent us from calling for memory that isn't there in many cases (asking for a 16g alloc when we've used up 24 on a 32g card will not fly). Any particular blockers on implementation or did the logic not pan out as you'd hoped?