Inference | KV prefix caching.

[lmcafee-nvidia]

What does this PR do ?

Implement KV cache prefix caching for dynamic batching inference, enabling requests with identical prompt prefixes to share computed KV cache blocks instead of recomputing them.

Key Features

• Block-level prefix sharing: Requests with matching prompt prefixes share KV cache blocks, reducing redundant computation
• Content-based hashing: Blocks are identified by hash of their token content for efficient prefix matching
• Request coordination: Multi-rank coordination prevents redundant KV computation when multiple requests match the same prefix blocks
• LRU eviction: Cached blocks use LRU eviction based on timestamps when memory is constrained
• CLI control: New --no-inference-dynamic-batching-enable-prefix-caching flag to disable when needed

Changes

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

• Add block hash tracking and content-based matching
• Implement hash-to-block-id mapping for prefix lookup
• Add reference counting for shared blocks
• LRU eviction for cached blocks

Dynamic Context (megatron/core/inference/contexts/dynamic_context.py):

• _find_matching_prefix_blocks() - Find cached blocks matching request prefix
• _compute_block_hashes() - Compute content hashes for blocks
• Block sharing logic in add_request()

Dynamic Engine (megatron/core/inference/engines/dynamic_engine.py):

• Request coordination to avoid redundant computation
• Pending block tracking during prefill
• Integration with chunked prefill scheduling

Inference Request (megatron/core/inference/inference_request.py):

• Add enable_prefix_caching flag
• Track matched prefix blocks per request

Arguments (megatron/training/arguments.py):

• Add --no-inference-dynamic-batching-enable-prefix-caching flag

How It Works

1. When a request arrives, compute hashes for each block of prompt tokens
2. Look up existing cached blocks with matching hashes
3. Reuse matched blocks (increment ref count) instead of allocating new ones
4. Only compute KV for unmatched (new) tokens
5. After prefill, cache newly computed blocks for future requests
6. Coordinate across DP ranks to prevent redundant computation

Test plan

• Unit tests for block hash computation and matching (57 tests in test_dynamic_prefix_caching.py)
• Unit tests for request coordination
• Tests for LRU eviction
• Tests for prefix caching enable/disable
• Tests for edge cases (empty prompts, single tokens, exact block boundaries)
• End-to-end inference test with prefix sharing

⚠️ 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.

[lmcafee-nvidia]

/ok to test f53d674

[lmcafee-nvidia]

/ok to test 5758bc3

[svcnvidia-nemo-ci]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/22584143291

[lmcafee-nvidia]

/ok to test 5a692b3

[lmcafee-nvidia]

/ok to test dc8fb69

[svcnvidia-nemo-ci]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/22598390284