From 053d5035cab43ae90c709209b8f88b57f93e40e8 Mon Sep 17 00:00:00 2001
From: Tairan Xu <34982739+Andrewxu313@users.noreply.github.com>
Date: Fri, 11 Sep 2026 15:52:32 +0100
Subject: [PATCH] build: declare fla-core (flash-linear-attention) as the
 Kimi-K3 KDA dependency

Kimi-linear/K3 hard-import fla (fla.modules, fla.ops.kda) for Kimi Delta Attention
but it was undeclared. Add fla-core>=0.5.0 to requirements.txt (feeds install_requires)
and document it in INSTALL.md and the K3 deploy guide.
---
 docs/INSTALL.md             | 1 +
 docs/deploy-kimi-k3-h200.md | 2 ++
 requirements.txt            | 1 +
 3 files changed, 4 insertions(+)

diff --git a/docs/INSTALL.md b/docs/INSTALL.md
index 1dba25f75..cccd94af2 100644
--- a/docs/INSTALL.md
+++ b/docs/INSTALL.md
@@ -9,6 +9,7 @@
   `libnuma-dev` on Debian/Ubuntu) — required for the `core_engine` JIT build
   (`#include <numa.h>`). `install_deps.sh` installs these automatically; without them
   the first server launch fails with `numa.h: No such file or directory`.
+- **Model-specific (Kimi-K3 / Kimi-linear)**: `fla-core>=0.5.0` (flash-linear-attention) provides the `fla` Kimi-Delta-Attention kernels. It is listed in `requirements.txt`, so `pip install .` (and `install_deps.sh`) pull it automatically — no separate step.
 - **GitHub access**: the repository is currently **private** — anonymous
   `git clone` and raw release-asset URLs fail (404). Authenticate first
   (e.g. `gh auth login`), clone via `gh repo clone batchgen-project/batchgen`,
diff --git a/docs/deploy-kimi-k3-h200.md b/docs/deploy-kimi-k3-h200.md
index e0ee8a9b6..76d435be3 100644
--- a/docs/deploy-kimi-k3-h200.md
+++ b/docs/deploy-kimi-k3-h200.md
@@ -51,6 +51,8 @@ and [Manual Installation](manual-installation.md). The default decode/prefill ex
 **NCCL**, which needs no extra build. (An optional DeepEP low-latency exchange is covered in
 §6.)
 
+Kimi-K3's KDA (Kimi Delta Attention) needs `fla-core>=0.5.0` (flash-linear-attention, the `fla` package). It is listed in `requirements.txt`, so the standard install pulls it automatically — no separate step.
+
 ## 4. Start the Server (2 × 8 H200)
 
 ### Mount shared memory
diff --git a/requirements.txt b/requirements.txt
index d951d7156..1961bd3fd 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -13,6 +13,7 @@ dill==0.3.7
 einops==0.8.1
 fastapi==0.121.1
 filelock==3.20.0
+fla-core>=0.5.0
 frozenlist==1.8.0
 fsspec==2023.10.0
 h11==0.16.0