From af7917c45a182572bb90a734542d091c71facd00 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Wed, 27 May 2026 12:44:44 +0400
Subject: [PATCH] Studio: don't kill chat-model llama-server when spawning
 helper backends

---
 studio/backend/core/inference/llama_cpp.py  | 17 +++++++++++++++--
 studio/backend/core/rag/captioner.py        |  7 ++++++-
 studio/backend/utils/datasets/llm_assist.py |  4 +++-
 3 files changed, 24 insertions(+), 4 deletions(-)

diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 0531dd0e93..b8d168ccc5 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -608,7 +608,19 @@ class LlamaCppBackend:
         3. unload_model() — terminates llama-server subprocess
     """
 
-    def __init__(self):
+    def __init__(self, kill_orphans: bool = True):
+        """Construct a backend wrapper around llama-server.
+
+        ``kill_orphans`` (default True): at construction time, reap any
+        llama-server processes lingering from a prior studio crash. Safe
+        for the global singleton (only one LlamaCppBackend exists at
+        startup). Pass ``False`` for short-lived secondary instances
+        spawned alongside an already-running chat-model server (e.g.
+        the RAG captioner helper, `_run_with_helper`) — otherwise the
+        constructor will kill the parent's healthy chat model because
+        it can't distinguish "another instance's healthy server" from
+        "a stale process".
+        """
         self._process: Optional[subprocess.Popen] = None
         self._port: Optional[int] = None
         self._model_identifier: Optional[str] = None
@@ -691,7 +703,8 @@ class LlamaCppBackend:
         # to decide whether to wait for the VRAM reclaim to finish.
         self._last_kill_monotonic: float = 0.0
 
-        self._kill_orphaned_servers()
+        if kill_orphans:
+            self._kill_orphaned_servers()
         atexit.register(self._cleanup)
 
     # ── Properties ────────────────────────────────────────────────
diff --git a/studio/backend/core/rag/captioner.py b/studio/backend/core/rag/captioner.py
index cc34e7c9a7..561dd5e659 100644
--- a/studio/backend/core/rag/captioner.py
+++ b/studio/backend/core/rag/captioner.py
@@ -74,7 +74,12 @@ def _load_helper_vlm() -> Optional[tuple[Any, str, str]]:
     try:
         from core.inference.llama_cpp import LlamaCppBackend
 
-        backend = LlamaCppBackend()
+        # kill_orphans=False is critical: the global singleton is
+        # already running the user's chat-model llama-server. Killing
+        # "orphans" here would reap that healthy chat process because
+        # the orphan-killer can't tell two LlamaCppBackend instances
+        # apart by PID ownership.
+        backend = LlamaCppBackend(kill_orphans = False)
         logger.info(
             "RAG captioner: loading helper VLM %s (%s) as fallback",
             _HELPER_REPO,
diff --git a/studio/backend/utils/datasets/llm_assist.py b/studio/backend/utils/datasets/llm_assist.py
index a36a121c5d..004dd5b354 100644
--- a/studio/backend/utils/datasets/llm_assist.py
+++ b/studio/backend/utils/datasets/llm_assist.py
@@ -155,7 +155,9 @@ def _run_with_helper(prompt: str, max_tokens: int = 256) -> Optional[str]:
     try:
         from core.inference.llama_cpp import LlamaCppBackend
 
-        backend = LlamaCppBackend()
+        # kill_orphans=False so the helper backend doesn't reap the
+        # parent's chat-model llama-server while loading itself.
+        backend = LlamaCppBackend(kill_orphans = False)
         logger.info(f"Loading helper model: {repo} ({variant})")
 
         ok = backend.load_model(