Fix/slow tier memory budget

.github/workflows/slow-tests.yml

@@ -30,6 +30,10 @@ jobs:
           df -h /
       - name: Add swap for fp32 CPU model boots
         run: |
+          # The runner image can ship /swapfile already active; fallocate on
+          # live swap fails with "Text file busy", so release it first.
+          sudo swapoff /swapfile 2>/dev/null || true
+          sudo rm -f /swapfile
           sudo fallocate -l 8G /swapfile
           sudo chmod 600 /swapfile
           sudo mkswap /swapfile

tests/integration/model_bridge/test_ssm_real_weights.py

@@ -34,13 +34,15 @@ class _Case:
     default: str  # canonical HF id, or "" for env-only (no small official checkpoint)
 
 
-# Ordered by how likely the checkpoint is small enough to actually run in CI.
+# Families with no CI-sized official checkpoint are env-only ("" default): an
+# fp32 CPU boot of granite-4.0-tiny-preview (~7B) alone peaks over 40 GB RSS,
+# which OOM-kills the 16 GB + 8 GB-swap nightly runner mid-file.
 CASES = [
     _Case("mamba1", "TL_MAMBA1_MODEL", "state-spaces/mamba-130m-hf"),
     _Case("mamba2", "TL_MAMBA2_MODEL", "AntonV/mamba2-130m-hf"),
-    _Case("granite", "TL_GRANITE_MODEL", "ibm-granite/granite-4.0-tiny-preview"),
-    _Case("nemotron_h", "TL_NEMOTRON_H_MODEL", "nvidia/NVIDIA-Nemotron-Nano-9B-v2"),
-    _Case("qwen3_next", "TL_QWEN3_NEXT_MODEL", "Qwen/Qwen3-Next-80B-A3B-Instruct"),
+    _Case("granite", "TL_GRANITE_MODEL", ""),
+    _Case("nemotron_h", "TL_NEMOTRON_H_MODEL", ""),
+    _Case("qwen3_next", "TL_QWEN3_NEXT_MODEL", ""),
     _Case("qwen3_5", "TL_QWEN3_5_MODEL", ""),
 ]