---
title: '49762'
pageLayout: default
slug: thoughts/inference/vllm-prs/49762
permalink: https://aarnphm.xyz/thoughts/inference/vllm-prs/49762.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
# [#49762](https://github.com/aarnphm/aarnphm.github.io/issues/49762) \[KV Connector\] Support NIXL P/D for hybrid MLA+SSM models

[vllm-project/vllm#49762](https://github.com/vllm-project/vllm/pull/49762) · njhill · MERGED · created 2026-07-24T20:46:32Z · merged 2026-07-29T03:50:31Z

Re-implements the intent of [vllm-project/vllm#44848](https://github.com/vllm-project/vllm/pull/44848) on top of the reworked NIXL connector, from the current architecture rather than the original patch. KimiLinear pools its KDA (GDN-typed MambaSpec) and MLA layers into shared HMA tensors, making every region dual-purpose. Since the mamba-page unification raises the attention block size until the MLA page equals the unified page, and FlashMLA fixes the kernel block at 64 tokens, both sides always exchange kernel-granularity geometry that is TP-independent for MLA: `block_size_ratio` stays 1 and TP variance is absorbed by the existing per-engine `physical_blocks_per_logical` handling. What remains are three real gaps:

- `register_kv_caches` marked a shared region’s FA view by whichever layer registered it first; a KDA-first ordering left dual-purpose regions flagged as head-sharded. Merge the MLA flag on the HMA dedup path so split handles replicate their FA descriptors.
- The push worker’s replicated-MLA WRITE branch asserted a single source rank, which hybrid MLA+SSM violates (sharded SSM targets every covered D rank). Replicate only the attention groups across the write ranks and route hybrid writes through the per-rank split handles, mirroring the pull side.
- Handshake validation skipped all FA block_len checks for mamba models; hybrid MLA can and now does enforce the TP-independent invariant that kernel-granularity block_lens match exactly.

The KDA conv state decomposes as GDN (equal Q/K/V sub-projections), so the existing 3-read conv transfer applies unchanged. Heterogeneous `physical_blocks_per_logical` still requires disabling prefix caching, as already enforced at handshake time.

Claude was used for this.

---

## comments

**mergify** (2026-07-28T00:27:56Z):

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @njhill.

<https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork>

