x

scripts/start_vllm.py

@@ -128,12 +128,14 @@ def launch_model(model_id, config, model_path, gpu_count):
     ctx = int(config.get("ctx", 8192))
     max_seqs = int(config.get("max_num_seqs", 64))
     gpu_util = float(config.get("gpu_util", 0.98))
+    served_model_name = config.get("served_model_name", model_id)
     
     log(f"Config: TP={tp_size}, Ctx={ctx}, Seqs={max_seqs}, Util={gpu_util}")
     
     # Build command
     cmd = [
         "vllm", "serve", model_path,
+        "--served-model-name", served_model_name,
         "--host", HOST,
         "--port", PORT,
         "--tensor-parallel-size", str(tp_size),
@@ -159,9 +161,16 @@ def launch_model(model_id, config, model_path, gpu_count):
     tool_call_parser = config.get("tool_call_parser")
     if tool_call_parser:
         cmd.extend(["--tool-call-parser", tool_call_parser])
-        cmd.extend(["--enable-auto-tool-choice"])
+        openclaw_compat = os.getenv("OPENCLAW_COMPAT", "false").lower() == "true"
+        enable_auto_tool_choice = config.get("enable_auto_tool_choice")
+        if enable_auto_tool_choice is None:
+            enable_auto_tool_choice = not openclaw_compat
+        if enable_auto_tool_choice:
+            cmd.extend(["--enable-auto-tool-choice"])
+            log("Added auto tool choice enabled")
+        else:
+            log("Auto tool choice disabled")
         log(f"Added tool call parser: {tool_call_parser}")
-        log("Added auto tool choice enabled")
     
     log(f"Command: {' '.join(cmd)}")