[fix] Fix missing fields in xqa kernel cache key #6282

lowsfer · 2025-07-23T07:01:30Z

Summary by CodeRabbit

New Features
- Enhanced runtime configuration to optionally account for position embedding type, improving flexibility in certain scenarios.
Bug Fixes
- Improved uniqueness and accuracy of runtime hash keys for better handling of model variations.
Refactor
- Updated internal logic to support the new optional parameter without affecting existing workflows.

coderabbitai · 2025-07-23T07:01:38Z

📝 Walkthrough

Walkthrough

The changes extend the XQAKernelRuntimeHashKey struct by adding an optional position_embedding_type field. Associated equality and hash functions are updated accordingly. The construction of this hash key is modified in all relevant places, including JIT and precompiled kernel implementations, to include the new optional field and adjust related parameters.

Changes

File(s)	Change Summary
Core common implementation `cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplCommon.cpp`	Modified `getRuntimeHashKeyFromXQAParams` to conditionally include `position_embedding_type` in the returned struct.
Core common header `cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplCommon.h`	Added optional `position_embedding_type` to `XQAKernelRuntimeHashKey`, updated equality operator and hash calculation.
JIT kernel implementation `cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpp`	Updated `getRuntimeHashKeyFromKernelMeta` to add the new optional field and modify `is_fp8_output` parameter type in hash key construction.
Precompiled kernel implementation `cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplPrecompiled.cpp`	Updated `XQAKernelRuntimeHashKey` construction in two locations to append `std::nullopt` for the new optional field and replace `0` with `false` for `is_fp8_output`.

Sequence Diagram(s)

sequenceDiagram
    participant Caller
    participant XQAParams
    participant HashKeyBuilder
    participant HashTable

    Caller->>HashKeyBuilder: getRuntimeHashKeyFromXQAParams(xqaParams, isXqaJit)
    HashKeyBuilder->>XQAParams: Access parameters
    alt isXqaJit == true
        HashKeyBuilder->>HashKeyBuilder: Set position_embedding_type = xqaParams.position_embedding_type
    else
        HashKeyBuilder->>HashKeyBuilder: Set position_embedding_type = nullopt
    end
    HashKeyBuilder->>HashTable: Use hash key (with optional position_embedding_type)

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Possibly related PRs

fix precompiled multi_query_token kernel not having is_fp8_out hash key #6279: Related to extending the XQAKernelRuntimeHashKey with additional optional fields, focusing on the is_fp8_output flag for precompiled kernels; complements this PR by addressing different fields within the same hashing mechanism.

Suggested reviewers

lucifer1004
kaiyux

Note

⚡️ Unit Test Generation is now available in beta!

Learn more here, or try it out under "Finishing Touches" below.

📜 Recent review details

Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between 2144ecd and 3be47f8.

📒 Files selected for processing (4)

cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplCommon.cpp (1 hunks)
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplCommon.h (2 hunks)
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpp (1 hunks)
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplPrecompiled.cpp (2 hunks)

🚧 Files skipped from review as they are similar to previous changes (4)

cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplJIT/decoderXQAImplJIT.cpp
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplPrecompiled.cpp
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplCommon.h
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplCommon.cpp

⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)

GitHub Check: Pre-commit Check

✨ Finishing Touches

📝 Generate Docstrings

🧪 Generate unit tests

Create PR with unit tests
Post copyable unit tests in a comment

🪧 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 generate unit tests to generate unit tests for 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 or @coderabbitai title anywhere in the PR title to generate the title automatically.

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.

lowsfer · 2025-07-23T07:35:08Z

/bot run

tensorrt-cicd · 2025-07-23T07:37:54Z

PR_Github #12672 [ run ] triggered by Bot

tensorrt-cicd · 2025-07-23T07:40:20Z

PR_Github #12673 [ run ] triggered by Bot

tensorrt-cicd · 2025-07-23T07:40:22Z

PR_Github #12672 [ run ] completed with state ABORTED

tensorrt-cicd · 2025-07-23T10:46:53Z

PR_Github #12673 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #9424 completed with status: 'FAILURE'

lowsfer · 2025-07-25T02:45:33Z

/bot run

tensorrt-cicd · 2025-07-25T02:54:27Z

PR_Github #12932 [ run ] triggered by Bot

tensorrt-cicd · 2025-07-25T05:19:54Z

PR_Github #12932 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #9642 completed with status: 'FAILURE'

Signed-off-by: Yao Yao <[email protected]>

lowsfer · 2025-07-28T06:36:15Z

/bot run

tensorrt-cicd · 2025-07-28T06:41:15Z

PR_Github #13160 [ run ] triggered by Bot

tensorrt-cicd · 2025-07-28T16:04:10Z

PR_Github #13160 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #9854 completed with status: 'FAILURE'

lowsfer · 2025-07-29T04:02:52Z

/bot run

tensorrt-cicd · 2025-07-29T04:08:24Z

PR_Github #13288 [ run ] triggered by Bot

tensorrt-cicd · 2025-07-29T11:20:13Z

PR_Github #13288 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #9926 completed with status: 'FAILURE'

cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplPrecompiled.cpp

lowsfer · 2025-07-31T08:37:58Z

/bot run

tensorrt-cicd · 2025-07-31T08:44:20Z

PR_Github #13640 [ run ] triggered by Bot

tensorrt-cicd · 2025-07-31T09:08:20Z

PR_Github #13640 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #10235 completed with status: 'SUCCESS'

Signed-off-by: Yao Yao <[email protected]> Signed-off-by: Lanyu Liao <[email protected]>

Signed-off-by: Yao Yao <[email protected]>

…VIDIA#6282 NVIDIA#6279 * [None][infra] Pin the version for triton to 3.3.1 (NVIDIA#6508) Signed-off-by: qqiao <[email protected]> * [None][infra] Pin the version for triton to 3.3.1 (NVIDIA#6508) (NVIDIA#6519) (NVIDIA#6549) Signed-off-by: Yanchao Lu <[email protected]> * [fix]: use safeInitRowMax instead of fp32_lowest to avoid NaN (NVIDIA#7087) Signed-off-by: Yao Yao <[email protected]> * [None][fix] Fix a numerical stability issue for XQA with spec dec Signed-off-by: Yao Yao <[email protected]> * fix typo Signed-off-by: Jhao-Ting Chen <[email protected]> * fix precompiled multi_query_token kernel not having is_fp8_out hash key (NVIDIA#6279) Signed-off-by: Jhao-Ting Chen <[email protected]> * [fix] Fix missing fields in xqa kernel cache key (NVIDIA#6282) Signed-off-by: Yao Yao <[email protected]> --------- Signed-off-by: qqiao <[email protected]> Signed-off-by: Yanchao Lu <[email protected]> Signed-off-by: Yao Yao <[email protected]> Signed-off-by: Jhao-Ting Chen <[email protected]> Co-authored-by: Emma Qiao <[email protected]> Co-authored-by: Yanchao Lu <[email protected]> Co-authored-by: Yao Yao <[email protected]>

lowsfer requested review from a team as code owners July 23, 2025 07:01

lowsfer requested review from Superjomn, dongxuy04 and symphonylyh July 23, 2025 07:01

lowsfer changed the base branch from main to release/0.21 July 23, 2025 07:01

lowsfer requested a review from a team as a code owner July 23, 2025 07:01

lowsfer force-pushed the fix-xqa-key branch from b7cb8db to 943bb3f Compare July 23, 2025 07:32

lowsfer changed the base branch from release/0.21 to main July 23, 2025 07:32

lowsfer force-pushed the fix-xqa-key branch from 943bb3f to 2144ecd Compare July 23, 2025 07:34

NVIDIA deleted a comment from tensorrt-cicd Jul 23, 2025

nv-yilinf mentioned this pull request Jul 23, 2025

[https://nvbugs/5398180][feat] Improve Llama4 performance for small max_seqlen cases #6306

Merged

[fix] include rope style in XQA kernel cache key

3be47f8

Signed-off-by: Yao Yao <[email protected]>

lowsfer force-pushed the fix-xqa-key branch from d36ffce to 3be47f8 Compare July 28, 2025 06:30

coderabbitai bot requested review from kaiyux and lucifer1004 July 28, 2025 06:31

symphonylyh requested a review from jhaotingc July 29, 2025 01:41

symphonylyh approved these changes Jul 29, 2025

View reviewed changes

cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderXQAImplPrecompiled.cpp Show resolved Hide resolved

lowsfer merged commit 942e080 into NVIDIA:main Aug 1, 2025
3 checks passed

lancelly pushed a commit to lancelly/TensorRT-LLM that referenced this pull request Aug 6, 2025

[fix] Fix missing fields in xqa kernel cache key (NVIDIA#6282)

c38aaad

Signed-off-by: Yao Yao <[email protected]> Signed-off-by: Lanyu Liao <[email protected]>

jain-ria pushed a commit to jain-ria/TensorRT-LLM that referenced this pull request Aug 7, 2025

[fix] Fix missing fields in xqa kernel cache key (NVIDIA#6282)

0b1ed2f

Signed-off-by: Yao Yao <[email protected]>

jhaotingc pushed a commit to jhaotingc/TensorRT-LLM that referenced this pull request Aug 25, 2025

[fix] Fix missing fields in xqa kernel cache key (NVIDIA#6282)

e0caeae

Signed-off-by: Yao Yao <[email protected]>

[fix] Fix missing fields in xqa kernel cache key #6282

[fix] Fix missing fields in xqa kernel cache key #6282

Uh oh!

Conversation

lowsfer commented Jul 23, 2025 • edited by coderabbitai bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary by CodeRabbit

Uh oh!

coderabbitai bot commented Jul 23, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Walkthrough

Changes

Sequence Diagram(s)

Estimated code review effort

Possibly related PRs

Suggested reviewers

Chat

Support

CodeRabbit Commands (Invoked using PR comments)

Other keywords and placeholders

Documentation and Community

Uh oh!

lowsfer commented Jul 23, 2025

Uh oh!

tensorrt-cicd commented Jul 23, 2025

Uh oh!

tensorrt-cicd commented Jul 23, 2025

Uh oh!

tensorrt-cicd commented Jul 23, 2025

Uh oh!

tensorrt-cicd commented Jul 23, 2025

Uh oh!

lowsfer commented Jul 25, 2025

Uh oh!

tensorrt-cicd commented Jul 25, 2025

Uh oh!

tensorrt-cicd commented Jul 25, 2025

Uh oh!

lowsfer commented Jul 28, 2025

Uh oh!

tensorrt-cicd commented Jul 28, 2025

Uh oh!

tensorrt-cicd commented Jul 28, 2025

Uh oh!

lowsfer commented Jul 29, 2025

Uh oh!

tensorrt-cicd commented Jul 29, 2025

Uh oh!

tensorrt-cicd commented Jul 29, 2025

Uh oh!

Uh oh!

lowsfer commented Jul 31, 2025

Uh oh!

tensorrt-cicd commented Jul 31, 2025

Uh oh!

tensorrt-cicd commented Jul 31, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

lowsfer commented Jul 23, 2025 •

edited by coderabbitai bot

Loading

coderabbitai bot commented Jul 23, 2025 •

edited

Loading