[Perf] Evaluate API: Using threads to infer signatures for eval batch runs instead of processes.

src/promptflow-devkit/promptflow/_proxy/_python_inspector_proxy.py

@@ -1,10 +1,11 @@
+import os
 import re
 from pathlib import Path
 from typing import Any, Dict, List
 
 from promptflow._constants import FlowEntryRegex
 from promptflow._core.entry_meta_generator import _generate_flow_meta
-from promptflow._sdk._constants import FLOW_META_JSON_GEN_TIMEOUT
+from promptflow._sdk._constants import FLOW_META_JSON_GEN_TIMEOUT, PF_FLOW_META_LOAD_IN_SUBPROCESS
 from promptflow._utils.flow_utils import resolve_python_entry_file
 
 from ._base_inspector_proxy import AbstractInspectorProxy
@@ -35,7 +36,7 @@ def get_entry_meta(
         **kwargs,
     ) -> Dict[str, Any]:
         timeout = kwargs.get("timeout", FLOW_META_JSON_GEN_TIMEOUT)
-        load_in_subprocess = kwargs.get("load_in_subprocess", True)
+        load_in_subprocess = os.environ.get(PF_FLOW_META_LOAD_IN_SUBPROCESS, "True").lower() == "true"
 
         flow_dag = {"entry": entry}
         # generate flow.json only for eager flow for now

src/promptflow-devkit/promptflow/_sdk/_constants.py

@@ -99,6 +99,7 @@ def _prepare_home_dir() -> Path:
 PF_SERVICE_DEBUG = "PF_SERVICE_DEBUG"
 PF_SYSTEM_METRICS_PREFIX = "__pf__"
 PF_FLOW_ENTRY_IN_TMP = "PF_FLOW_ENTRY_IN_TMP"
+PF_FLOW_META_LOAD_IN_SUBPROCESS = "PF_FLOW_META_LOAD_IN_SUBPROCESS"
 
 LOCAL_MGMT_DB_PATH = (HOME_PROMPT_FLOW_DIR / "pf.sqlite").resolve()
 LOCAL_MGMT_DB_SESSION_ACQUIRE_LOCK_PATH = (HOME_PROMPT_FLOW_DIR / "pf.sqlite.lock").resolve()

src/promptflow-evals/promptflow/evals/evaluate/_batch_run_client/batch_run_context.py

@@ -3,7 +3,7 @@
 # ---------------------------------------------------------
 import os
 
-from promptflow._sdk._constants import PF_FLOW_ENTRY_IN_TMP
+from promptflow._sdk._constants import PF_FLOW_ENTRY_IN_TMP, PF_FLOW_META_LOAD_IN_SUBPROCESS
 from promptflow._utils.user_agent_utils import ClientUserAgentUtil
 from promptflow.tracing._integrations._openai_injector import inject_openai_api, recover_openai_api
 
@@ -23,10 +23,12 @@ def __enter__(self):
 
         if isinstance(self.client, ProxyClient):
             os.environ[PF_FLOW_ENTRY_IN_TMP] = "true"
+            os.environ[PF_FLOW_META_LOAD_IN_SUBPROCESS] = "false"
 
     def __exit__(self, exc_type, exc_val, exc_tb):
         if isinstance(self.client, CodeClient):
             recover_openai_api()
 
         if isinstance(self.client, ProxyClient):
             os.environ.pop(PF_FLOW_ENTRY_IN_TMP, None)
+            os.environ.pop(PF_FLOW_META_LOAD_IN_SUBPROCESS, None)