From 0783813fdfec255c2f66a0a700e7d883efed4739 Mon Sep 17 00:00:00 2001
From: Alison Shao <54658187+alisonshao@users.noreply.github.com>
Date: Thu, 4 Jun 2026 01:25:00 -0700
Subject: [PATCH] ci: cache HF hub for base-a-test-cpu to avoid Hub 429 flakes
 (#27092)

---
 .github/workflows/pr-test.yml | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/.github/workflows/pr-test.yml b/.github/workflows/pr-test.yml
index 18bed3548..1631aa39a 100644
--- a/.github/workflows/pr-test.yml
+++ b/.github/workflows/pr-test.yml
@@ -267,6 +267,8 @@ jobs:
       (needs.check-changes.outputs.main_package == 'true')
     runs-on: ubuntu-latest
     timeout-minutes: 240
+    env:
+      HF_HOME: ${{ github.workspace }}/.hf-cache
     strategy:
       fail-fast: false
       max-parallel: ${{ fromJson(needs.check-changes.outputs.partitions)['base-a-test-cpu'].max_parallel }}
@@ -316,6 +318,16 @@ jobs:
         run: |
           uv pip install -e "python[dev]" --index-strategy unsafe-best-match --prerelease allow
 
+      # Hosted runners are ephemeral, so models are re-fetched every run and the
+      # Hub occasionally returns 429s. Persist the HF cache in GitHub's cache
+      # storage (rolling key + restore-keys) so warm runs never hit the network.
+      - name: Cache HF hub
+        uses: actions/cache@v4
+        with:
+          path: ${{ github.workspace }}/.hf-cache
+          key: hf-cpu-${{ matrix.partition }}-${{ github.run_id }}
+          restore-keys: hf-cpu-${{ matrix.partition }}-
+
       - name: Run test
         timeout-minutes: 10
         env: