[Feature][MLA] Support Kimi K3 attention by maoxx241 · Pull Request #14839 · vllm-project/vllm-ascend

maoxx241 · 2026-08-24T09:34:06Z

What this PR does / why we need it?

Adds Kimi K3 target and DSpark MLA execution on Ascend.
Preserves per-group RoPE/no-RoPE semantics and causal/non-causal draft metadata through the MLA backend.
Adds Kimi K3 rotary inputs and MLA weight post-processing, including native floating-point MLAPO handling on A5.
Covers MLA metadata, mixed rotary modes, graph updates, and rotary behavior with focused unit tests.

Dependencies

[Feature][Ops] Add Kimi K3 AscendC operators #14426 provides the required AscendC operator bindings.
The Kimi K3 model adapter PR consumes this attention implementation.

How was this patch tested?

Python bytecode compilation for all changed Python files.
Ruff 0.14.0 format and lint checks for all changed Python files.
Focused coverage in tests/ut/attention/a2/test_mla_v1.py and tests/ut/ops/test_rotary_embedding.py.

Does this PR introduce any user-facing change?

Yes. It enables Kimi K3 MLA execution on Ascend.

vLLM version: v0.27.1
vLLM main: vllm-project/vllm@ba07e4a

github-actions · 2026-08-24T09:35:43Z

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

A PR should do only one thing, smaller PRs enable faster reviews.
Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

Tip

💡 Consider Linking a Related Issue or RFC

Your PR title contains the [Feature] tag, indicating a bug fix or new feature.

Linking a related issue or RFC in the PR description is strongly encouraged — it gives reviewers helpful context and speeds up the review. You can use any of these keywords:

Fixes #<issue_number>
Closes #<issue_number>
Resolves #<issue_number>
Refs #<rfc_or_issue_number> (for RFCs)

🙏 Thanks for helping us keep the project well-organized!

gemini-code-assist · 2026-08-24T09:37:08Z

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces support for Kimi K3 Multi-Head Latent Attention (MLA) execution on Ascend hardware. The changes include updates to the MLA backend to handle rotary-positioning and weight post-processing, as well as the propagation of causal and non-causal draft metadata. These enhancements enable the vLLM Ascend implementation to support Kimi K3 models effectively.

Highlights

Kimi K3 Support: Implemented Kimi K3 MLA execution on Ascend, including specific rotary-position handling and weight post-processing.
Metadata Integration: Integrated causal and non-causal draft metadata support within the MLA backend to ensure correct execution flow.
Testing: Added comprehensive unit tests covering MLA metadata and rotary behavior in the Ascend environment.

New Features

🧠 You can now enable Memory (public preview) to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize the Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counterproductive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

Suggested PR Title:

[Attention][Feature] Support MLA without RoPE, output gating, and non-causal speculative decoding

Suggested PR Summary:

### What this PR does / why we need it?
This pull request introduces support for Multi-Head Latent Attention (MLA) without Rotary Position Embeddings (RoPE), adds output gating (`g_proj`) support, and enables non-causal (bidirectional) speculative decoding. Specifically, it adds `get_identity_cos_and_sin_mla` to skip rotary cache when RoPE is disabled, handles non-causal sparse modes in Flash Attention, and updates the MLAPO prolog to support native floating-point weights on A5.

Feedback:
Four high-severity issues were identified in the review:
1. Potential `KeyError` in `AscendMLAMetadataBuilder` due to `.attn` suffix mismatch when looking up `static_forward_context`.
2. A similar `KeyError` in `AscendMLAImpl` when `fa_quant_layer` is enabled.
3. A dtype mismatch error in `mla_preprocess_only_decode` when `use_mla_rope` is `False` due to hardcoded `torch.bfloat16`.
4. A potential memory corruption/runtime failure in `_exec_kv_no_rope` due to passing a non-contiguous `k_pe` tensor to `reshape_and_cache`.

### Does this PR introduce _any_ user-facing change?
No, these are internal backend optimizations and feature alignments for speculative decoding and MLA execution on Ascend devices.

### How was this patch tested?
Tested via new unit tests in `tests/ut/attention/a2/test_mla_v1.py` covering draft layer rope mode, target layer nope mode, and mixed rope modes, as well as `tests/ut/ops/test_rotary_embedding.py` for identity cos/sin cache skipping.

maoxx241 · 2026-08-24T11:16:07Z

Implementation walkthrough

MLA metadata and causality

vllm_ascend/attention/mla_v1.py carries two model-level properties into the runtime metadata:

whether the layer uses MLA RoPE;
whether the attention block is causal.

Kimi target MLA uses the normal causal path. Kimi MLA DSpark uses a non-causal multi-token draft block, so its metadata keeps causal=False and avoids a triangular FIA mask that would silently change the draft computation.

No-RoPE Kimi path

For layers without MLA RoPE, get_identity_cos_and_sin_mla() returns stable-shape identity buffers. The same metadata and graph interfaces can therefore be reused without adding a second attention backend. A2 routes no-RoPE decode through generic preprocessing; A5 passes empty RoPE tensors to the CANN MLA Prolog V3 contract.

Output gate

When the upstream Kimi module provides g_proj, the Ascend attention implementation applies the Kimi output gate after attention and before the output projection. Models without this projection keep the existing MLA behavior.

A5 MLA Prolog V3 weight preparation

The post-loading path prepares the fused Kimi projections for npu_mla_prolog_v3:

split weight_dq and weight_dkv_kr from the fused Q/KV input projection;
prepare weight_uq_qr and pad head dimensions when required by the operator;
keep native floating-point weights in mode 0;
keep W8A8/MXFP8 weights and their scales in mode 3;
use the prepared W_UK layout and the layer's effective padded head count.

This uses the standard model post-loading hook and the existing CANN operator interface.

Rotary and SP handling

vllm_ascend/ops/rotary_embedding.py keeps position handling compatible with sequence-parallel execution and provides the identity MLA buffers. vllm_ascend/ops/mla.py passes the draft causality, output gate, and RoPE mode from the upstream model module into the backend.

Coverage

The tests cover causal/non-causal metadata, no-RoPE identity buffers, Kimi output gating, A5 native and quantized Prolog V3 preparation, padded heads, and rotary/SP position behavior.

github-actions · 2026-08-26T11:53:20Z

This pull request has conflicts, please resolve those before we can evaluate the pull request.

Preserve per-group RoPE semantics, non-causal multi-token decode metadata, explicit no-RoPE execution, and A5 MLA preprocessing for Kimi K3 target and draft layers. Reuse the current device MLA implementation and PCP compatibility paths. Signed-off-by: maoxx241 <maomaoyu870@gmail.com>

Keep absent cos/sin metadata intact before calling the existing no-RoPE query and KV paths. The PCP prefill changes introduced unconditional slicing before those paths could bypass rotation. Preserve the distinct actual-query and padded-KV lengths for RoPE layers, and cover pure and mixed prefill with numeric Q/K/V and cache-write assertions. Signed-off-by: maoxx241 <maomaoyu870@gmail.com>

### What this PR does / why we need it? This integration PR enables Kimi K3 text, multimodal, MTP, DSpark, Prefix Cache, and P/D serving on Ascend. The implementation is reviewed through the atomic PRs below; this parent owns the Kimi K3 deployment and validation guide. ### Recommended merge order | Order | PR | Responsibility | | ---: | --- | --- | | 1 | #14426 | AscendC KDA, chunk gated delta rule, Dequant-SiTU, and MX-SiTU operators | | 2 | #14597 | Hybrid Mamba state-copy, asynchronous accepted-token snapshots, and Ascend launch-grid correctness | | 3 | #14598 | KDA attention execution and fused RMSNorm gate | | 4 | #14839 | MLA attention and rotary execution | | 5 | #14840 | Attention-residual Triton fusion | | 6 | #14599 | SiTU MoE, shared-expert execution, and K3 ModelSlim quantization adaptation | | 7 | #14600 | Text, multimodal, MTP, and DSpark model registration; ViT FIA contiguous inputs; model-owned QuaRot shared-layer conversion | | 8 | #14601 | DSpark speculative-decoding runtime and generic shared-layer hook integration | | 9 | #14765 | TP8/TP16 GQA/MLA DSpark KV grouping, speculative capacity, and per-rank DCP table sizing | | 10 | #14602 | Hybrid P/D transfer, proxy retry, and graph-safe stateful handoffs | After these PRs merge, the parent-owned change is: - `docs/source/tutorials/models/Kimi-K3.md` - `docs/source/tutorials/models/index.md` The guide covers reduced and full checkpoints, single-node TP16, four-node DP4/TP16/EP64, GQA/MLA DSpark, two-node P/D, QuaRot, Prefix Cache, server-side validation, GPQA, and performance reporting. ### How was this patch tested? - Fused norm-gate dispatch: 39 targeted CPU tests pass (4 existing skips) and 22 real A3 NPU numerical cases pass on each of vLLM v0.27.1 and the pinned upstream revision. Coverage includes FP16/BF16, sigmoid/SiLU, packed gate strides, residual/prenorm, and input preservation. The actual upstream CustomOp resolves to the Ascend fused implementation; three ACLGraph replays with fresh inputs match upstream native results. CI mypy and `bash format.sh ci` pass. A5 performance and full-model serving were not rerun for this change. - State and capacity regressions: 94 targeted CPU tests pass, with two post-v0.27.1 coordinator-API cases skipped. Coverage includes real InputBatch replacement/reordering, accepted-token ownership across scheduling modes, GDN metadata, Mamba copy ordering, scheduler/worker capacity agreement, and writes to the final speculative Mamba slots at DCP1/DCP4. `bash format.sh ci` passes. Full-model NPU serving was not rerun for the snapshot/capacity changes. - Handoff graph selection: 12 CPU dispatch cases and 25 GDN metadata tests pass. Distributed NPU end-to-end validation was not rerun for this graph-selection change. - All changed Python files pass syntax compilation. - The parent includes the current child implementations plus the deployment documentation. - Focused coverage includes KDA, MLA, SiTU MoE, Mamba state copy, DSpark, compressed/hybrid KV cache, Prefix Cache, one-token P/D handoff, Mooncake transfer, and model registration. - Full-checkpoint integration coverage includes text, multimodal, tools, streaming, QuaRot, C64/C128, TP8/TP16, two-node P/D, four-node GQA DSpark, the known K3 accuracy/hang cases, and GPQA-Diamond. Detailed accuracy and performance results remain in the PR comments. ### Does this PR introduce any user-facing change? Yes. Kimi K3 can be deployed with text, multimodal, MTP, DSpark, Prefix Cache, and P/D serving on Ascend. - vLLM version: v0.27.1 - vLLM main: vllm-project/vllm@ba07e4a --------- Signed-off-by: maoxx241 <maomaoyu870@gmail.com> Signed-off-by: weinachuan <weinachuan1@huawei.com> Signed-off-by: zongersama <48584200+zongersama@users.noreply.github.com> Signed-off-by: yolic66 <747731294@qq.com> Signed-off-by: Dawn952 <zhaojunbo13@huawei.com> Signed-off-by: MQ <maomaoyu870@gmail.com> Co-authored-by: weinachuan <weinachuan1@huawei.com> Co-authored-by: zongersama <48584200+zongersama@users.noreply.github.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Co-authored-by: yolic66 <747731294@qq.com> Co-authored-by: Dawn952 <zhaojunbo13@huawei.com>

maoxx241 · 2026-08-28T01:37:29Z

Closing this child PR following the merge of parent #14454.

### What this PR does / why we need it? This integration PR enables Kimi K3 text, multimodal, MTP, DSpark, Prefix Cache, and P/D serving on Ascend. The implementation is reviewed through the atomic PRs below; this parent owns the Kimi K3 deployment and validation guide. ### Recommended merge order | Order | PR | Responsibility | | ---: | --- | --- | | 1 | vllm-project#14426 | AscendC KDA, chunk gated delta rule, Dequant-SiTU, and MX-SiTU operators | | 2 | vllm-project#14597 | Hybrid Mamba state-copy, asynchronous accepted-token snapshots, and Ascend launch-grid correctness | | 3 | vllm-project#14598 | KDA attention execution and fused RMSNorm gate | | 4 | vllm-project#14839 | MLA attention and rotary execution | | 5 | vllm-project#14840 | Attention-residual Triton fusion | | 6 | vllm-project#14599 | SiTU MoE, shared-expert execution, and K3 ModelSlim quantization adaptation | | 7 | vllm-project#14600 | Text, multimodal, MTP, and DSpark model registration; ViT FIA contiguous inputs; model-owned QuaRot shared-layer conversion | | 8 | vllm-project#14601 | DSpark speculative-decoding runtime and generic shared-layer hook integration | | 9 | vllm-project#14765 | TP8/TP16 GQA/MLA DSpark KV grouping, speculative capacity, and per-rank DCP table sizing | | 10 | vllm-project#14602 | Hybrid P/D transfer, proxy retry, and graph-safe stateful handoffs | After these PRs merge, the parent-owned change is: - `docs/source/tutorials/models/Kimi-K3.md` - `docs/source/tutorials/models/index.md` The guide covers reduced and full checkpoints, single-node TP16, four-node DP4/TP16/EP64, GQA/MLA DSpark, two-node P/D, QuaRot, Prefix Cache, server-side validation, GPQA, and performance reporting. ### How was this patch tested? - Fused norm-gate dispatch: 39 targeted CPU tests pass (4 existing skips) and 22 real A3 NPU numerical cases pass on each of vLLM v0.27.1 and the pinned upstream revision. Coverage includes FP16/BF16, sigmoid/SiLU, packed gate strides, residual/prenorm, and input preservation. The actual upstream CustomOp resolves to the Ascend fused implementation; three ACLGraph replays with fresh inputs match upstream native results. CI mypy and `bash format.sh ci` pass. A5 performance and full-model serving were not rerun for this change. - State and capacity regressions: 94 targeted CPU tests pass, with two post-v0.27.1 coordinator-API cases skipped. Coverage includes real InputBatch replacement/reordering, accepted-token ownership across scheduling modes, GDN metadata, Mamba copy ordering, scheduler/worker capacity agreement, and writes to the final speculative Mamba slots at DCP1/DCP4. `bash format.sh ci` passes. Full-model NPU serving was not rerun for the snapshot/capacity changes. - Handoff graph selection: 12 CPU dispatch cases and 25 GDN metadata tests pass. Distributed NPU end-to-end validation was not rerun for this graph-selection change. - All changed Python files pass syntax compilation. - The parent includes the current child implementations plus the deployment documentation. - Focused coverage includes KDA, MLA, SiTU MoE, Mamba state copy, DSpark, compressed/hybrid KV cache, Prefix Cache, one-token P/D handoff, Mooncake transfer, and model registration. - Full-checkpoint integration coverage includes text, multimodal, tools, streaming, QuaRot, C64/C128, TP8/TP16, two-node P/D, four-node GQA DSpark, the known K3 accuracy/hang cases, and GPQA-Diamond. Detailed accuracy and performance results remain in the PR comments. ### Does this PR introduce any user-facing change? Yes. Kimi K3 can be deployed with text, multimodal, MTP, DSpark, Prefix Cache, and P/D serving on Ascend. - vLLM version: v0.27.1 - vLLM main: vllm-project/vllm@ba07e4a --------- Signed-off-by: maoxx241 <maomaoyu870@gmail.com> Signed-off-by: weinachuan <weinachuan1@huawei.com> Signed-off-by: zongersama <48584200+zongersama@users.noreply.github.com> Signed-off-by: yolic66 <747731294@qq.com> Signed-off-by: Dawn952 <zhaojunbo13@huawei.com> Signed-off-by: MQ <maomaoyu870@gmail.com> Co-authored-by: weinachuan <weinachuan1@huawei.com> Co-authored-by: zongersama <48584200+zongersama@users.noreply.github.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Co-authored-by: yolic66 <747731294@qq.com> Co-authored-by: Dawn952 <zhaojunbo13@huawei.com> Signed-off-by: d30086105 <denghaojie1@h-partners.com>

### What this PR does / why we need it? This integration PR enables Kimi K3 text, multimodal, MTP, DSpark, Prefix Cache, and P/D serving on Ascend. The implementation is reviewed through the atomic PRs below; this parent owns the Kimi K3 deployment and validation guide. ### Recommended merge order | Order | PR | Responsibility | | ---: | --- | --- | | 1 | vllm-project#14426 | AscendC KDA, chunk gated delta rule, Dequant-SiTU, and MX-SiTU operators | | 2 | vllm-project#14597 | Hybrid Mamba state-copy, asynchronous accepted-token snapshots, and Ascend launch-grid correctness | | 3 | vllm-project#14598 | KDA attention execution and fused RMSNorm gate | | 4 | vllm-project#14839 | MLA attention and rotary execution | | 5 | vllm-project#14840 | Attention-residual Triton fusion | | 6 | vllm-project#14599 | SiTU MoE, shared-expert execution, and K3 ModelSlim quantization adaptation | | 7 | vllm-project#14600 | Text, multimodal, MTP, and DSpark model registration; ViT FIA contiguous inputs; model-owned QuaRot shared-layer conversion | | 8 | vllm-project#14601 | DSpark speculative-decoding runtime and generic shared-layer hook integration | | 9 | vllm-project#14765 | TP8/TP16 GQA/MLA DSpark KV grouping, speculative capacity, and per-rank DCP table sizing | | 10 | vllm-project#14602 | Hybrid P/D transfer, proxy retry, and graph-safe stateful handoffs | After these PRs merge, the parent-owned change is: - `docs/source/tutorials/models/Kimi-K3.md` - `docs/source/tutorials/models/index.md` The guide covers reduced and full checkpoints, single-node TP16, four-node DP4/TP16/EP64, GQA/MLA DSpark, two-node P/D, QuaRot, Prefix Cache, server-side validation, GPQA, and performance reporting. ### How was this patch tested? - Fused norm-gate dispatch: 39 targeted CPU tests pass (4 existing skips) and 22 real A3 NPU numerical cases pass on each of vLLM v0.27.1 and the pinned upstream revision. Coverage includes FP16/BF16, sigmoid/SiLU, packed gate strides, residual/prenorm, and input preservation. The actual upstream CustomOp resolves to the Ascend fused implementation; three ACLGraph replays with fresh inputs match upstream native results. CI mypy and `bash format.sh ci` pass. A5 performance and full-model serving were not rerun for this change. - State and capacity regressions: 94 targeted CPU tests pass, with two post-v0.27.1 coordinator-API cases skipped. Coverage includes real InputBatch replacement/reordering, accepted-token ownership across scheduling modes, GDN metadata, Mamba copy ordering, scheduler/worker capacity agreement, and writes to the final speculative Mamba slots at DCP1/DCP4. `bash format.sh ci` passes. Full-model NPU serving was not rerun for the snapshot/capacity changes. - Handoff graph selection: 12 CPU dispatch cases and 25 GDN metadata tests pass. Distributed NPU end-to-end validation was not rerun for this graph-selection change. - All changed Python files pass syntax compilation. - The parent includes the current child implementations plus the deployment documentation. - Focused coverage includes KDA, MLA, SiTU MoE, Mamba state copy, DSpark, compressed/hybrid KV cache, Prefix Cache, one-token P/D handoff, Mooncake transfer, and model registration. - Full-checkpoint integration coverage includes text, multimodal, tools, streaming, QuaRot, C64/C128, TP8/TP16, two-node P/D, four-node GQA DSpark, the known K3 accuracy/hang cases, and GPQA-Diamond. Detailed accuracy and performance results remain in the PR comments. ### Does this PR introduce any user-facing change? Yes. Kimi K3 can be deployed with text, multimodal, MTP, DSpark, Prefix Cache, and P/D serving on Ascend. - vLLM version: v0.27.1 - vLLM main: vllm-project/vllm@ba07e4a --------- Signed-off-by: maoxx241 <maomaoyu870@gmail.com> Signed-off-by: weinachuan <weinachuan1@huawei.com> Signed-off-by: zongersama <48584200+zongersama@users.noreply.github.com> Signed-off-by: yolic66 <747731294@qq.com> Signed-off-by: Dawn952 <zhaojunbo13@huawei.com> Signed-off-by: MQ <maomaoyu870@gmail.com> Co-authored-by: weinachuan <weinachuan1@huawei.com> Co-authored-by: zongersama <48584200+zongersama@users.noreply.github.com> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com> Co-authored-by: yolic66 <747731294@qq.com> Co-authored-by: Dawn952 <zhaojunbo13@huawei.com>

maoxx241 requested review from realliujiaxu, weijinqian0 and zzzzwwjj as code owners August 24, 2026 09:34

github-actions Bot added module:tests module:ops labels Aug 24, 2026

This was referenced Aug 24, 2026

[Feature][Model] Register Kimi K3 model adapters #14600

Closed

[Feature][Spec Decode] Add Kimi K3 DSpark execution #14601

Closed

maoxx241 mentioned this pull request Aug 24, 2026

[Feature][Model] Add Kimi K3 support for vLLM 0.27 #14454

Merged

gemini-code-assist Bot reviewed Aug 24, 2026

View reviewed changes

Comment thread vllm_ascend/attention/mla_v1.py Outdated

Comment thread vllm_ascend/attention/mla_v1.py Outdated

Comment thread vllm_ascend/device/device_op.py Outdated

Comment thread vllm_ascend/attention/mla_v1.py

maoxx241 force-pushed the codex/kimi-k3-mla-main branch 5 times, most recently from 19f5288 to aa7c434 Compare August 25, 2026 09:44

github-actions Bot added the merge-conflicts label Aug 26, 2026

maoxx241 added 2 commits August 27, 2026 23:21

maoxx241 force-pushed the codex/kimi-k3-mla-main branch from aa7c434 to 4d8c0a7 Compare August 27, 2026 15:21

github-actions Bot removed the merge-conflicts label Aug 27, 2026

maoxx241 closed this Aug 28, 2026

Navigation Menu

Repository navigation

[Feature][MLA] Support Kimi K3 attention - #14839

[Feature][MLA] Support Kimi K3 attention#14839
maoxx241 wants to merge 2 commits into
vllm-project:mainfrom
maoxx241:codex/kimi-k3-mla-main

maoxx241 commented Aug 24, 2026 •

edited by github-actions Bot

Loading

Uh oh!

github-actions Bot commented Aug 24, 2026

💡 Consider Linking a Related Issue or RFC

Uh oh!

gemini-code-assist Bot commented Aug 24, 2026

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

maoxx241 commented Aug 24, 2026

Uh oh!

github-actions Bot commented Aug 26, 2026

Uh oh!

maoxx241 commented Aug 28, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

maoxx241 commented Aug 24, 2026 • edited by github-actions Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What this PR does / why we need it?

Dependencies

How was this patch tested?

Does this PR introduce any user-facing change?

Uh oh!

github-actions Bot commented Aug 24, 2026

💡 Consider Linking a Related Issue or RFC

Uh oh!

gemini-code-assist Bot commented Aug 24, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

maoxx241 commented Aug 24, 2026

Implementation walkthrough

MLA metadata and causality

No-RoPE Kimi path

Output gate

A5 MLA Prolog V3 weight preparation

Rotary and SP handling

Coverage

Uh oh!

github-actions Bot commented Aug 26, 2026

Uh oh!

maoxx241 commented Aug 28, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

maoxx241 commented Aug 24, 2026 •

edited by github-actions Bot

Loading