chore(release): source signed HTP bundle from qcom-ai-hub/geniex LFS

Replace the anonymous S3 GET in overlay-htp with a sparse LFS
checkout of qcom-ai-hub/geniex, so China eng can commit signed
bundles via ordinary PR instead of pushing to S3 through the
geo-barrier. Requires secret QCOM_AI_HUB_GENIEX_READ_TOKEN.

Update notes/release.md and the self-signed release-notes fragment
in .github/scripts/release.js to reference the LFS path.

Refs qcom-ai-hub/geniex#1309, qcom-ai-hub/geniex#1297.

Signed-off-by: Mengsheng Wu <mengshen@qti.qualcomm.com>

Author: mengshengwu

.github/scripts/release.js

@@ -34,7 +34,7 @@ module.exports = async ({ github, context, core }) => {
   const htpNote =
     HTP_SIGNED === "true"
       ? `## Hexagon HTP\n\nMicrosoft-signed HTP catalog (llama.cpp @ \`${LLAMA_SHA}\`). No cert import required on Windows on Snapdragon.`
-      : `## Hexagon HTP\n\nSelf-signed HTP catalog (llama.cpp @ \`${LLAMA_SHA}\`). End users must enable test signing and import \`ggml-htp-v1.cer\` per [notes/run.md](../blob/${VERSION}/notes/run.md).\n\nOperators: ship \`libggml-htp-to-sign-${LLAMA_SHA}.zip\` to the signing pipeline, then upload the signed result as \`libggml-htp-${LLAMA_SHA}.zip\` to \`s3://qaihub-public-assets/llama-cpp/\` and re-run this release.`;
+      : `## Hexagon HTP\n\nSelf-signed HTP catalog (llama.cpp @ \`${LLAMA_SHA}\`). End users must enable test signing and import \`ggml-htp-v1.cer\` per [notes/run.md](../blob/${VERSION}/notes/run.md).\n\nOperators: ship \`libggml-htp-to-sign-${LLAMA_SHA}.zip\` to the signing pipeline, then commit the signed result as \`sdk/signed-htp/libggml-htp-${LLAMA_SHA}.zip\` to \`qcom-ai-hub/geniex\` via PR (LFS-tracked) and re-run this release.`;
 
   let release;
   for await (const res of github.paginate.iterator(
@@ -46,7 +46,7 @@ module.exports = async ({ github, context, core }) => {
   }
 
   // Merge htpNote into the release body, replacing any stale HTP section from a
-  // prior run (e.g. self-signed → Microsoft-signed once the S3 bundle lands).
+  // prior run (e.g. self-signed → Microsoft-signed once the signed bundle lands).
   const mergeHtpNote = (body) => {
     const base = (body || "").replace(/## Hexagon HTP[\s\S]*$/m, "").trimEnd();
     return base ? `${base}\n\n${htpNote}` : htpNote;

.github/workflows/release.yml

@@ -101,8 +101,9 @@ jobs:
       version_tag: ${{ needs.resolve-tag.outputs.tag }}
       upload_artifact: true
 
-  # Overlay the Microsoft-signed HTP bundle from S3 if available; otherwise
-  # the Windows SDK stays self-signed and gets marked as such downstream.
+  # Overlay the Microsoft-signed HTP bundle from qcom-ai-hub/geniex LFS if
+  # available; otherwise the Windows SDK stays self-signed and gets marked
+  # as such downstream.
   overlay-htp:
     name: overlay-htp (windows-arm64)
     runs-on: ubuntu-latest
@@ -120,24 +121,35 @@ jobs:
       - name: Init llama.cpp submodule
         run: git submodule update --init --depth=1 third-party/llama.cpp
 
+      - name: Fetch signed HTP archive from qcom-ai-hub/geniex
+        uses: actions/checkout@v7
+        with:
+          repository: qcom-ai-hub/geniex
+          ref: main
+          lfs: true
+          path: signed-htp-src
+          sparse-checkout: sdk/signed-htp
+          sparse-checkout-cone-mode: false
+          token: ${{ secrets.QCOM_AI_HUB_GENIEX_READ_TOKEN }}
+          fetch-depth: 1
+
       - uses: actions/download-artifact@v8
         with:
           name: sdk-windows-arm64
           path: sdk-windows-arm64
 
-      - name: Overlay Microsoft-signed HTP bundle from S3 (if available)
+      - name: Overlay Microsoft-signed HTP bundle from LFS (if available)
         id: htp
         run: |
           set -euo pipefail
           sha=$(git -C third-party/llama.cpp rev-parse --short=6 HEAD)
-          url="https://qaihub-public-assets.s3.us-west-2.amazonaws.com/llama-cpp/libggml-htp-${sha}.zip"
-          if curl -fsSL -o htp-signed.zip "$url"; then
-            unzip -o htp-signed.zip -d sdk-windows-arm64/lib/llama_cpp/
+          src="signed-htp-src/sdk/signed-htp/libggml-htp-${sha}.zip"
+          if [[ -f "$src" ]]; then
+            unzip -o "$src" -d sdk-windows-arm64/lib/llama_cpp/
             signed=true
           else
             signed=false
           fi
-          rm -f htp-signed.zip
           echo "Signed HTP bundle for llama.cpp @ ${sha}: ${signed}"
           echo "signed=${signed}" >> "$GITHUB_OUTPUT"
           echo "llama_sha=${sha}" >> "$GITHUB_OUTPUT"

notes/release.md

@@ -165,22 +165,24 @@ The per-tag manifest is byte-stable across workflow re-runs of the same tag —
 
 ## Hexagon HTP signing
 
-The Windows ARM64 SDK ships `libggml-htp.cat` plus `libggml-htp-v{68,69,73,75,79,81}.so` — Windows refuses to load them unsigned. Release CI runs an `overlay-htp` job **before** `build-cli` that `curl`s `s3://qaihub-public-assets/llama-cpp/libggml-htp-<sha>.zip`, where `<sha>` is the `third-party/llama.cpp` short SHA. Both the installer and the SDK zip end up with the same HTP files:
+The Windows ARM64 SDK ships `libggml-htp.cat` plus `libggml-htp-v{68,69,73,75,79,81}.so` — Windows refuses to load them unsigned. Release CI runs an `overlay-htp` job **before** `build-cli` that sparse-checks-out `sdk/signed-htp/libggml-htp-<sha>.zip` from `qcom-ai-hub/geniex` (LFS-tracked), where `<sha>` is the `third-party/llama.cpp` short SHA. Both the installer and the SDK zip end up with the same HTP files:
 
 - **Hit** — overlay the Microsoft-signed files into the SDK artifact; `build-cli` packages them into the installer; release normally.
 - **Miss** — keep the self-signed build. The SDK name gets a `-selfsigned` suffix, and the release also carries `ggml-htp-v1.cer` (users import it) and `libggml-htp-to-sign-<sha>.zip` (operators submit it for signing).
 
-The S3 bundle must contain exactly these eight files at the zip root: `libggml-htp.cat`, `libggml-htp.inf`, and `libggml-htp-v{68,69,73,75,79,81}.so`.
+The signed bundle must contain exactly these eight files at the zip root: `libggml-htp.cat`, `libggml-htp.inf`, and `libggml-htp-v{68,69,73,75,79,81}.so`.
+
+The cross-repo checkout uses `secrets.QCOM_AI_HUB_GENIEX_READ_TOKEN` — a GitHub App / fine-grained PAT scoped to `qcom-ai-hub/geniex` with `contents: read` + `metadata: read`. If CI reports `signed=false` but the bundle is merged on `main`, first check that this secret has not expired.
 
 ### Promoting self-signed → Microsoft-signed
 
 1. Download `libggml-htp-to-sign-<sha>.zip` from the draft release.
 2. Submit for Microsoft signing.
    a. Put the `.cat` `.inf` and all `.so` files into `ATT\libggml-htp\` in samba;
    b. Submit Jenkins pipeline, fill path with `\path\to\ATT`, other field use default or first param.
-   c. Get singed files from `ATT\Glymur\01000\ExtractedDrivers`.
-   d. keep files in a zip with the same files (without `.inf`) at the root.
-3. Upload the result to `s3://qaihub-public-assets/llama-cpp/libggml-htp-<sha>.zip`, with public acl.
+   c. Get signed files from `ATT\Glymur\01000\ExtractedDrivers`.
+   d. Repack the signed files (without `.inf`) into a zip with the same layout at the root.
+3. Commit the result to `qcom-ai-hub/geniex` at `sdk/signed-htp/libggml-htp-<sha>.zip` — `git lfs install` locally, add the zip on a branch, open a PR titled per [CONTRIBUTING.md](../CONTRIBUTING.md) (for example `chore(release): add signed HTP bundle for llama.cpp <sha>`), and get a maintainer to squash-merge into `main`.
 4. Re-run the Release workflow for the same tag.
 
 ## Windows installer signing gate