[CI] Size the CPU stage from the live partition model (#33329)

.github/workflows/pr-test.yml

@@ -334,13 +334,26 @@ jobs:
           key: hf-cpu-${{ matrix.partition }}-${{ github.run_id }}
           restore-keys: hf-cpu-${{ matrix.partition }}-
 
+      # Pinned SHA so every shard splits against the same snapshot; without
+      # the file run_suite falls back to the (drifting) in-source est_time.
+      - name: Fetch live partition model
+        if: needs.check-changes.outputs.partition_model_sha != ''
+        run: |
+          rm -f /tmp/partition-model.json
+          URL="https://raw.githubusercontent.com/sgl-project/sglang-ci-stats/${{ needs.check-changes.outputs.partition_model_sha }}/model.json"
+          curl --fail --silent --show-error --max-time 15 --retry 3 --retry-delay 2 \
+            "$URL" -o /tmp/partition-model.json
+
+      # compute_partitions reads this back as the per-shard budget, so it
+      # drives the fanout rather than capping it -- shrinking it buys more
+      # shards, not shorter ones.
       - name: Run test
-        timeout-minutes: 10
+        timeout-minutes: 15
         env:
           CONTINUE_ON_ERROR_FLAG: ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
         run: |
           cd test/
-          python3 run_suite.py --hw cpu --suite base-a-test-cpu --auto-partition-id ${{ matrix.partition }} --auto-partition-size ${{ fromJson(needs.check-changes.outputs.partitions)['base-a-test-cpu'].size }} $CONTINUE_ON_ERROR_FLAG
+          python3 run_suite.py --hw cpu --suite base-a-test-cpu --auto-partition-id ${{ matrix.partition }} --auto-partition-size ${{ fromJson(needs.check-changes.outputs.partitions)['base-a-test-cpu'].size }} --partition-model-file /tmp/partition-model.json $CONTINUE_ON_ERROR_FLAG
 
   # Runs on 5090 (32GB, SM120)
   base-b-test-1-gpu-small: