@@ -0,0 +1,60 @@
+name: Diagnose recurrent Linux abort
+
+on:
+  workflow_dispatch:
+  push:
+    branches:
+      - test/diagnose-linux-recurrent-abort
+
+jobs:
+  diagnose:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - name: native-repack
+            cmake-args: -DGGML_NATIVE=ON -DGGML_CPU_REPACK=ON
+          - name: generic-repack
+            cmake-args: -DGGML_NATIVE=OFF -DGGML_CPU_REPACK=ON
+          - name: native-no-repack
+            cmake-args: -DGGML_NATIVE=ON -DGGML_CPU_REPACK=OFF
+    name: ${{ matrix.name }}
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          submodules: recursive
+
+      - name: Set up Python
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.11"
+
+      - name: Show runner CPU
+        run: lscpu
+
+      - name: Install debugger
+        run: |
+          sudo apt-get update
+          sudo apt-get install --yes gdb
+
+      - name: Build package
+        env:
+          CMAKE_ARGS: ${{ matrix.cmake-args }}
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install uv
+          python -m uv pip install -e '.[all]' --verbose
+
+      - name: Reproduce with native diagnostics
+        continue-on-error: true
+        run: python tests/diagnose_recurrent_context.py
+
+      - name: Capture native backtrace
+        continue-on-error: true
+        run: |
+          gdb --batch \
+            -ex "set pagination off" \
+            -ex run \
+            -ex "thread apply all bt" \
+            --args python tests/diagnose_recurrent_context.py
@@ -0,0 +1,32 @@
+import hashlib
+import multiprocessing
+
+from huggingface_hub import hf_hub_download
+
+import llama_cpp
+
+
+model_path = hf_hub_download(
+    repo_id="QuantFactory/mamba-130m-hf-GGUF",
+    filename="mamba-130m-hf.Q2_K.gguf",
+)
+
+with open(model_path, "rb") as model_file:
+    model_sha256 = hashlib.file_digest(model_file, "sha256").hexdigest()
+
+print(f"model_path={model_path}", flush=True)
+print(f"model_sha256={model_sha256}", flush=True)
+print(llama_cpp.llama_print_system_info().decode(), flush=True)
+
+model = llama_cpp.Llama(
+    model_path,
+    n_ctx=32,
+    n_batch=32,
+    n_ubatch=32,
+    n_threads=multiprocessing.cpu_count(),
+    n_threads_batch=multiprocessing.cpu_count(),
+    logits_all=False,
+    verbose=True,
+)
+print("context construction succeeded", flush=True)
+model.close()