[CI] Optimize test execution with single-GPU parallelism and log collection

Author: EmmonsCurse

fastdeploy/model_executor/layers/batch_invariant_ops/batch_invariant_ops.py

@@ -690,15 +690,13 @@ def addmm_batch_invariant(
     return result
 
 
-# def _log_softmax_batch_invariant(x: paddle.Tensor, axis: int = -1, out=None) -> paddle.Tensor:
-#     result = log_softmax(input=x, axis=axis)
-#     # Handle out parameter if provided
-#     if out is not None:
-#         out.copy_(result)
-#         return out
-#     return result
-def _log_softmax_batch_invariant(x: paddle.Tensor, axis: int = -1) -> paddle.Tensor:
-    return log_softmax(input=x, axis=axis)
+def _log_softmax_batch_invariant(x: paddle.Tensor, axis: int = -1, out=None) -> paddle.Tensor:
+    result = log_softmax(input=x, axis=axis)
+    # Handle out parameter if provided
+    if out is not None:
+        out.copy_(result)
+        return out
+    return result
 
 
 def mean_batch_invariant(

scripts/coverage_run.sh

@@ -327,6 +327,19 @@ if [ "$failed_count" -ne 0 ]; then
     echo "Failed test cases are listed in $failed_tests_file"
     cat "$failed_tests_file"
 
+    # clean the empty directories
+    if [ -d "${run_path}/unittest_logs" ]; then
+        echo "Cleaning empty directories..."
+
+        # perform multi-round clean until no more empty directories are found
+        while true; do
+            before=$(find "${run_path}/unittest_logs" -type d | wc -l)
+            find "${run_path}/unittest_logs" -mindepth 1 -type d -empty -delete || true
+            after=$(find "${run_path}/unittest_logs" -type d | wc -l)
+            [ "$before" -eq "$after" ] && break
+        done
+    fi
+
     # Only package logs when there are failures
     echo "===================================="
     echo "Step 5: Packaging logs (only on failure)"

tests/e2e/test_EB_Lite_serving.py

@@ -1389,9 +1389,8 @@ def test_streaming_chat_finish_reason(openai_client):
 
 def test_profile_reset_block_num():
     """测试profile reset_block_num功能，与baseline diff不能超过5%"""
-    log_file = "./log/config.log"
-    # log_dir = os.getenv("FD_LOG_DIR", "log")
-    # log_file = os.path.join(log_dir, "config.log")
+    log_dir = os.getenv("FD_LOG_DIR", "log")
+    log_file = os.path.join(log_dir, "config.log")
     baseline = 31446
 
     if not os.path.exists(log_file):

tests/e2e/test_ernie_03b_pd_router_v1_ipc.py

@@ -418,4 +418,4 @@ def test_non_chat_usage_non_stream(api_url):
     total_tokens = usage["completion_tokens"] + usage["prompt_tokens"]
     assert payload["max_tokens"] >= usage["completion_tokens"], "completion_tokens大于max_tokens"
     assert payload["metadata"]["min_tokens"] <= usage["completion_tokens"], "completion_tokens小于min_tokens"
-    assert usage["total_tokens"] != total_tokens, "total_tokens不等于prompt_tokens + completion_tokens"
+    assert usage["total_tokens"] == total_tokens, "total_tokens不等于prompt_tokens + completion_tokens"

tests/e2e/test_ernie_03b_pd_router_v1_rdma_global_cache.py

@@ -218,7 +218,7 @@ def setup_and_run_server():
     for k, v in mooncake_env.items():
         env_prefill[k] = v
 
-    prefill_log_path = os.path.join(prefill_log_dir, "nohup")
+    prefill_log_path = os.path.join(prefill_log_dir, "nohup.log")
     prefill_cmd = [
         sys.executable,
         "-m",