From 67f01ee7e0f0adec29ad0119c5632c5821b1f8cf Mon Sep 17 00:00:00 2001
From: WSHAPER <42714629+WSHAPER@users.noreply.github.com>
Date: Mon, 19 Oct 2026 15:06:05 +0200
Subject: [PATCH] fix: adapt llama-server to current xllamacpp binding

xllamacpp 2026.10.11549 renames CommonParams.hostname to hostnames
(now a list) and drops use_mmap/use_mlock, so the setattr loop crashed
at startup and took down every consumer CI. SERVER_KEYS updated, the
config sends hostnames as a list, and keys the binding no longer knows
are skipped with a warning instead of killing the server.

Verified against 2026.10.11549: server boots on a known-good model and
serves /health and /v1/chat/completions; unknown keys log and continue.

Assisted-by: opencode:zai/glm-5.3
Signed-off-by: WSHAPER <42714629+WSHAPER@users.noreply.github.com>
---
 lib/llama_server.py    | 11 +++++++----
 lib/task_processors.py |  2 +-
 2 files changed, 8 insertions(+), 5 deletions(-)

diff --git a/lib/llama_server.py b/lib/llama_server.py
index 7513703..3c18ddf 100644
--- a/lib/llama_server.py
+++ b/lib/llama_server.py
@@ -21,7 +21,6 @@
     "n_gpu_layers", "main_gpu", "split_mode", "tensor_split",
     "no_kv_offload", "no_op_offload", "flash_attn_type",
     # Memory
-    "use_mmap", "use_mlock",
     "cache_type_k", "cache_type_v",
     "cache_prompt", "cache_idle_slots", "cache_ram_mib",
     # RoPE / YaRN
@@ -32,7 +31,7 @@
     "reasoning_format", "enable_reasoning",
     "use_jinja", "chat_template", "enable_chat_template",
     # Server
-    "hostname", "port", "n_threads_http",
+    "hostnames", "port", "n_threads_http",
     "timeout_read", "timeout_write",
     "api_keys", "api_prefix",
     # Misc
@@ -49,8 +48,12 @@ def main() -> None:
     if cfg.get("mmproj_path"):
         p.mmproj.path = cfg["mmproj_path"]
     for k in SERVER_KEYS:
-        if k in cfg:
-            setattr(p, k, cfg[k])
+        if k not in cfg:
+            continue
+        if not hasattr(p, k):
+            print(f"llama_server: ignoring server key '{k}' (not supported by this xllamacpp)", file=sys.stderr)
+            continue
+        setattr(p, k, cfg[k])
 
     server = xlc.Server(p)  # noqa: F841 — held to keep the C++ server thread alive
     while True:
diff --git a/lib/task_processors.py b/lib/task_processors.py
index 2ea17ad..35502f3 100644
--- a/lib/task_processors.py
+++ b/lib/task_processors.py
@@ -150,7 +150,7 @@ def generate_chat_model(file_name: str) -> ChatOpenAI:
 
     server_cfg: dict = {
         "model_path": path,
-        "hostname": "127.0.0.1",
+        "hostnames": ["127.0.0.1"],
         "port": port,
         "n_gpu_layers": n_gpu_layers,
         "n_batch": loader_config.get("n_batch", 512),