ci: isolate recurrent verbosity abort

Author: abetlen

.github/workflows/diagnose-recurrent-linux.yaml

@@ -42,55 +42,31 @@ jobs:
           python -m pip install uv
           python -m uv pip install -e '.[all]' --verbose
 
-      - name: Recurrent test alone
+      - name: Construct recurrent context with verbose disabled
         continue-on-error: true
-        run: |
-          python -m pytest -vv -s \
-            tests/test_llama.py::test_recurrent_model_prompt_cache_reset
-
-      - name: Recurrent test after low-level model
-        continue-on-error: true
-        run: |
-          python -m pytest -vv -s \
-            tests/test_llama.py::test_real_model \
-            tests/test_llama.py::test_recurrent_model_prompt_cache_reset
+        env:
+          VERBOSE: "0"
+        run: python tests/diagnose_recurrent_context.py
 
-      - name: Recurrent test after high-level model
+      - name: Construct recurrent context with verbose enabled
         continue-on-error: true
-        run: |
-          python -m pytest -vv -s \
-            tests/test_llama.py::test_real_llama \
-            tests/test_llama.py::test_recurrent_model_prompt_cache_reset
-
-      - name: Recurrent test after repeated-prompt model
-        continue-on-error: true
-        run: |
-          python -m pytest -vv -s \
-            tests/test_llama.py::test_real_llama_repeated_prompt_cache \
-            tests/test_llama.py::test_recurrent_model_prompt_cache_reset
+        env:
+          VERBOSE: "1"
+        run: python tests/diagnose_recurrent_context.py
 
-      - name: Reproduce original test sequence
+      - name: Run original recurrent test
         continue-on-error: true
         run: |
           python -m pytest -vv -s \
-            tests/test_llama.py::test_llama_cpp_version \
-            tests/test_llama.py::test_llama_cpp_tokenization \
-            tests/test_llama.py::test_real_model \
-            tests/test_llama.py::test_real_llama \
-            tests/test_llama.py::test_real_llama_repeated_prompt_cache \
             tests/test_llama.py::test_recurrent_model_prompt_cache_reset
 
-      - name: Capture original-sequence backtrace
+      - name: Capture verbose-disabled backtrace
         continue-on-error: true
+        env:
+          VERBOSE: "0"
         run: |
           gdb --batch \
             -ex "set pagination off" \
             -ex run \
             -ex "thread apply all bt" \
-            --args python -m pytest -vv -s \
-              tests/test_llama.py::test_llama_cpp_version \
-              tests/test_llama.py::test_llama_cpp_tokenization \
-              tests/test_llama.py::test_real_model \
-              tests/test_llama.py::test_real_llama \
-              tests/test_llama.py::test_real_llama_repeated_prompt_cache \
-              tests/test_llama.py::test_recurrent_model_prompt_cache_reset
+            --args python tests/diagnose_recurrent_context.py

tests/diagnose_recurrent_context.py

@@ -1,5 +1,6 @@
 import hashlib
 import multiprocessing
+import os
 
 from huggingface_hub import hf_hub_download
 
@@ -18,6 +19,9 @@
 print(f"model_sha256={model_sha256}", flush=True)
 print(llama_cpp.llama_print_system_info().decode(), flush=True)
 
+verbose = os.environ.get("VERBOSE") == "1"
+print(f"verbose={verbose}", flush=True)
+
 model = llama_cpp.Llama(
     model_path,
     n_ctx=32,
@@ -26,7 +30,7 @@
     n_threads=multiprocessing.cpu_count(),
     n_threads_batch=multiprocessing.cpu_count(),
     logits_all=False,
-    verbose=True,
+    verbose=verbose,
 )
 print("context construction succeeded", flush=True)
 model.close()

tests/test_llama.py

@@ -309,7 +309,7 @@ def _assert_prompt_cache_reset_handles_history_edit(
         n_threads=multiprocessing.cpu_count(),
         n_threads_batch=multiprocessing.cpu_count(),
         logits_all=False,
-        verbose=True,
+        verbose=False,
     )
 
     assert model._is_recurrent is is_recurrent