Inference | Hybrid prefix caching.

[lmcafee-nvidia]

What does this PR do ?

Add Mamba state prefix caching for hybrid Transformer-Mamba models, enabling KV cache prefix sharing to also share corresponding Mamba conv/SSM states.

Key Features

• Mamba state caching: When KV cache blocks are reused due to shared prefixes, the corresponding Mamba states can now also be cached and restored, avoiding redundant recomputation
• Memory-budgeted cache: New --inference-dynamic-batching-prefix-caching-mamba-gb argument controls the memory budget for cached Mamba states
• LRU eviction: Mamba cache uses LRU eviction based on KV block timestamps when the cache is full
• Automatic invalidation: When KV blocks are evicted, their associated Mamba states are automatically invalidated

Changes

Core Implementation (megatron/core/inference/contexts/dynamic_context.py, megatron/core/inference/engines/dynamic_engine.py):

• Fixed-size tensor pool for cached Mamba conv/SSM states
• Block-to-slot mapping for associating KV blocks with Mamba cache slots
• Store Mamba state at block boundaries after prefill completes
• Restore cached Mamba state when prefix matches during request scheduling

Block Allocator (megatron/core/inference/contexts/dynamic_block_allocator.py):

• Invalidate Mamba state when KV blocks are evicted

Arguments (megatron/training/arguments.py):

• Add --inference-dynamic-batching-prefix-caching-mamba-gb parameter
• Auto-enable chunked prefill when Mamba prefix caching is used

Test plan

• Unit tests for cache allocation/deallocation
• Unit tests for LRU eviction (20 tests in tests/unit_tests/inference/contexts/test_mamba_prefix_caching.py)
• Tests for memory budget edge cases
• Tests for state store/restore integration
• End-to-end inference test with hybrid model

⚠️ For major changes (either in lines of code or in its impact), please make sure to first share a design doc with the team. If you're unsure what's the best way to do so, contact the @mcore-oncall.

Contribution process

flowchart LR
    A[Pre-checks] --> B[PR Tests]
    subgraph Code Review/Approval
        C1[Expert Review] --> C2[Final Review]
    end
    B --> C1
    C2 --> D[Merge]

Loading

Pre-checks

• I want this PR in a versioned release and have added the appropriate Milestone (e.g., Core 0.8)
• I have added relevant unit tests
• I have added relevant functional tests
• I have added proper typing to my code Typing guidelines
• I have added relevant documentation
• I have run the autoformatter.sh on my PR

Code review

The following process is enforced via the CODEOWNERS file for changes into megatron/core. For changes outside of megatron/core, it is up to the PR author whether or not to tag the Final Reviewer team.

For MRs into `main` branch

Feel free to message or comment the @mcore-oncall to help accelerate your merge into main. The less complex your PR is, the faster it will be approved and merged!

(Step 1): Add PR label Expert Review

(Step 2): Collect the expert reviewers reviews

1. Attach the Expert Review label when your PR is ready for review.
2. GitHub auto-assigns expert reviewers based on your changes. They will get notified and pick up your PR soon.

⚠️ Only proceed to the next step once all reviewers have approved, merge-conflict are resolved and the CI is passing.
Final Review might get declined if these requirements are not fulfilled.

(Step 3): Final Review

1. Add Final Review label
2. GitHub auto-assigns final reviewers based on your changes. They will get notified and pick up your PR soon.

(Optional Step 4): Cherry-pick into release branch

If this PR also needs to be merged into core_r* release branches, after this PR has been merged, select Cherry-pick to open a new PR into the release branch.

For MRs into `dev` branch

The proposed review process for `dev` branch is under active discussion.

MRs are mergable after one approval by either eharper@nvidia.com or zijiey@nvidia.com.

Merging your PR

Any member of core-adlr and core-nemo will be able to merge your PR.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.