From 504a723c5ce3dc82ce63e439a92324f1503e8624 Mon Sep 17 00:00:00 2001
From: IC <ic.lomugdang@gmail.com>
Date: Sun, 27 Sep 2026 16:06:10 +0200
Subject: [PATCH] QVAC-25280 feat[api]: bump qvac-fabric to 10549.4.0 in
 @qvac/fabric 0.17.1

Moves packages/fabric, the only vcpkg consumer of the qvac-fabric port, from
10549.3.0 to 10549.4.0 (Ternary Bonsai support on CPU, Metal, CUDA and
Vulkan) and releases @qvac/fabric 0.17.1.

0.17.1 also ships the Windows backend packaging from #4520, which landed
after 0.17.0 without a changelog entry: the win32-x64 prebuild now carries its
ggml backend modules, which 0.17.0 omitted.

The addons on the npm @qvac/fabric runtime take this through their existing
^0.17.0 range, so none of them is bumped here.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 packages/fabric/CHANGELOG.md | 22 ++++++++++++++++++++++
 packages/fabric/package.json |  2 +-
 packages/fabric/vcpkg.json   |  2 +-
 3 files changed, 24 insertions(+), 2 deletions(-)

diff --git a/packages/fabric/CHANGELOG.md b/packages/fabric/CHANGELOG.md
index 5355279dd28..f681d31245a 100644
--- a/packages/fabric/CHANGELOG.md
+++ b/packages/fabric/CHANGELOG.md
@@ -1,5 +1,27 @@
 # Changelog
 
+## [0.17.1] - 2026-09-26
+
+### Changed
+
+- `qvac-fabric` dependency bumped `10549.3.0` -> `10549.4.0`:
+  - Fixed a scheduler re-reserve on every causal-attention toggle.
+  - Fixed WebGPU normalization and copy dispatches that exceeded the
+    dispatch limit on large tensors.
+  - Added Ternary Bonsai model support on CPU, Metal, CUDA and Vulkan.
+  - Added chunked GDN prefill for CUDA/HIP (CUTLASS tensorized MMA) and a
+    SIMD-group GDN prefill for Metal.
+  - Added fused BF16 NVFP4 MLP prefill for CUDA.
+
+### Fixed
+
+- Windows builds now ship the ggml backend modules
+  (`prebuilds/win32-x64/qvac__fabric/*.dll`) next to the runtime. Since the
+  `10549.3.0` port, Windows builds the CPU variants and GPU backends as
+  runtime-loaded modules, and `0.17.0` staged them only for Android and Linux,
+  so a Windows consumer registered no backend and every model load failed with
+  `no backends are loaded`.
+
 ## [0.17.0] - 2026-09-23
 
 ### Changed
diff --git a/packages/fabric/package.json b/packages/fabric/package.json
index 747a4fa9a5c..4941409c712 100644
--- a/packages/fabric/package.json
+++ b/packages/fabric/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@qvac/fabric",
-  "version": "0.17.0",
+  "version": "0.17.1",
   "description": "Shared bare addon hosting the qvac-fabric (forked llama.cpp + ggml) runtime for QVAC inference addons",
   "addon": true,
   "engines": {
diff --git a/packages/fabric/vcpkg.json b/packages/fabric/vcpkg.json
index 1acd2eff1a8..b202bea1537 100644
--- a/packages/fabric/vcpkg.json
+++ b/packages/fabric/vcpkg.json
@@ -6,7 +6,7 @@
     },
     {
       "name": "qvac-fabric",
-      "version>=": "10549.3.0",
+      "version>=": "10549.4.0",
       "features": [
         "hip-backend",
         "vector-index"