---
title: Kimi K3 vLLM PR reading order
tags:
  - inference
  - vllm
  - new-architectures
pageLayout: default
slug: thoughts/inference/kimi-k3-prs
permalink: https://aarnphm.xyz/thoughts/inference/kimi-k3-prs.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
# Kimi K3 vLLM PR reading order

Snapshot: vLLM main @ 2026-10-11.

## 4. PR reading order

| #  | PR                                                                                                                                                                                                              | title                                                        | author                                   | date        | what to look at                                                                                                                                   |
| -- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------ | ---------------------------------------- | ----------- | ------------------------------------------------------------------------------------------------------------------------------------------------- |
| 1  | [#50089](https://github.com/aarnphm/aarnphm.github.io/issues/50089)                                                                                                                                             | Add Kimi K3 support: model files and kernels \[1/N\]         | Jiangyun Zhu                             | 07-29       | The bulk: `nvidia/{kda,kda_metadata,mla,model}.py`, `fused_kda_decode_kernel.cu`, MLA key-concat kernel, FlashKDA cmake, custom AG/RS collectives |
| 2  | [#50090](https://github.com/aarnphm/aarnphm.github.io/issues/50090)                                                                                                                                             | Add AttnRes kernels                                          | Thien Tran                               | 07-28       | `attn_res_kernel.cu` + Triton ref: the block/prefix math in §1.4                                                                                  |
| 3  | [#50000](https://github.com/aarnphm/aarnphm.github.io/issues/50000)                                                                                                                                             | \[New model\] Kimi K3                                        | Jiangyun Zhu                             | 07-30       | Glue: config, `kimi_gdn_linear_attn.py` rewrite, partial-prefix-cache tests, DSpark MLA, mxfp4/trtllm MoE changes                                 |
| 4  | [#51855](https://github.com/aarnphm/aarnphm.github.io/issues/51855)                                                                                                                                             | \[K3\] support recoverssm for K3                             | Jiangyun Zhu                             | 08-17       | `ops/recoverssm.py` verify/commit, `use_kda_recoverssm` resolution, extra state shapes                                                            |
| 5  | [#53614](https://github.com/aarnphm/aarnphm.github.io/issues/53614)                                                                                                                                             | Internal prefix checkpoints w/ partial prefix caching + spec | Ziming Huang                             | 09-06       | `get_mamba_prefill_checkpoint_position`, `is_mamba_prefill_checkpoint_valid`, reserved block, `replace_existing_hashes`                           |
| 6  | [#56794](https://github.com/aarnphm/aarnphm.github.io/issues/56794)                                                                                                                                             | Keep transient checkpoints out of prefix-cache eviction      | Sy03                                     | 09-14       | Why a reused checkpoint slot must evict stale hashes                                                                                              |
| 7  | [#55747](https://github.com/aarnphm/aarnphm.github.io/issues/55747)                                                                                                                                             | Fix `assert 0 <= checkpoint_idx < len(blocks)`               | Wentao Ye                                | 09-07       | Pin the reserved index at allocation time, not at cache time                                                                                      |
| 8  | [#51483](https://github.com/aarnphm/aarnphm.github.io/issues/51483)                                                                                                                                             | Do not classify a stateless first chunk as a decode          | sashko-zakharchuk                        | 09-16       | 1-token first chunk + recycled slot; `no_prior_state`                                                                                             |
| 9  | [#54255](https://github.com/aarnphm/aarnphm.github.io/issues/54255)                                                                                                                                             | FlashInfer speculative KDA backend                           | Duncan Moss                              | 09-30       | Packed fused spec decode; support matrix                                                                                                          |
| 10 | [#52988](https://github.com/aarnphm/aarnphm.github.io/issues/52988)                                                                                                                                             | Variable-length decode for K3 adaptive verification          | Albert Cheng                             | 09-22       | `AttentionCGSupport.ALWAYS`, device-side offsets, confidence head                                                                                 |
| 11 | [#50484](https://github.com/aarnphm/aarnphm.github.io/issues/50484)                                                                                                                                             | \[Kimi-K3\] DCP support                                      | Summer Yang                              | 08-10       | direct A2A LSE reduce; `MLADCPManager`                                                                                                            |
| 12 | [#50493](https://github.com/aarnphm/aarnphm.github.io/issues/50493)                                                                                                                                             | DCP partial prefix cache hit                                 | Summer Yang                              | 08-17       | coordinator/block_table under DCP                                                                                                                 |
| 13 | [#59229](https://github.com/aarnphm/aarnphm.github.io/issues/59229)                                                                                                                                             | Test prefix cache reuse with KV offload, P/D and DCP         | Jiangyun Zhu                             | 10-02       | The integration contract (§3.4)                                                                                                                   |
| 14 | [#50487](https://github.com/aarnphm/aarnphm.github.io/issues/50487)                                                                                                                                             | Tap pre-norm AttnRes mixture as DFlash aux state             | Rahul Chalamala                          | 08-14       | What “hidden state” means under AttnRes                                                                                                           |
| 15 | [#50912](https://github.com/aarnphm/aarnphm.github.io/issues/50912) / [#50656](https://github.com/aarnphm/aarnphm.github.io/issues/50656) / [#56875](https://github.com/aarnphm/aarnphm.github.io/issues/56875) | Shard shared expert (non-mega, SP, DeepEP v2)                | Wentao Ye / Tyler Smith / Elvir Crnčević | 08-04…10-08 | SP contract: AG → partial → RS                                                                                                                    |
| 16 | [#56159](https://github.com/aarnphm/aarnphm.github.io/issues/56159)                                                                                                                                             | Avoid KDA mixed-batch gather/scatter (+5–8% E2E)             | Wentao Ye                                | 09-10       | `spec_token_start` contiguous-slice fast path                                                                                                     |
| 17 | [#53396](https://github.com/aarnphm/aarnphm.github.io/issues/53396)                                                                                                                                             | DS conv-state layout in fused KDA decode                     | Canlin Guo                               | 08-27       | Layout coupling to P/D                                                                                                                            |
| 18 | [#59733](https://github.com/aarnphm/aarnphm.github.io/issues/59733)                                                                                                                                             | Declare max_tp_shards on the DSpark MLA KV spec              | Hyukjoon Lee                             | 10-08       | Spec-equality → KV grouping                                                                                                                       |

Also useful:

- Frontends: [#50104](https://github.com/aarnphm/aarnphm.github.io/issues/50104) (Rust), [#50093](https://github.com/aarnphm/aarnphm.github.io/issues/50093) (Python), [#58358](https://github.com/aarnphm/aarnphm.github.io/issues/58358) token-aware marker parsing.
- Startup and robustness: [#54634](https://github.com/aarnphm/aarnphm.github.io/issues/54634), [#55774](https://github.com/aarnphm/aarnphm.github.io/issues/55774) (RecoverSSM startup).
- Perf: [#54015](https://github.com/aarnphm/aarnphm.github.io/issues/54015) (gate merge), [#54261](https://github.com/aarnphm/aarnphm.github.io/issues/54261) (native AttnRes), [#51725](https://github.com/aarnphm/aarnphm.github.io/issues/51725) (adaptive spec token budget, \~60% DSpark TTFT).

---

