x

app/model_catalog.py

@@ -100,6 +100,8 @@ def resolve_model_profile(
         "selected_model": model_key,
         "model_name": _resolve_profile_model_path(profile, model_root, model_key),
         "served_model_name": profile.get("served_model_name", model_key),
+        "dtype": _to_str(profile.get("dtype")),
+        "quantization": _to_str(profile.get("quantization")),
         "max_model_len": _to_int(profile.get("ctx"), 8192),
         "max_num_seqs": _to_int(profile.get("max_num_seqs"), 64),
         "max_tokens": _to_int(profile.get("max_tokens"), 4096),

app/start_openai.py

@@ -19,7 +19,8 @@ def build_command() -> list[str]:
     host = str(runtime["openai_host"])
     port = str(runtime["openai_port"])
     api_key = runtime["api_key"] or ""
-    dtype = str(runtime["dtype"])
+    dtype = str(updates["dtype"] or runtime["dtype"])
+    quantization = str(updates["quantization"] or "").strip()
     revision = runtime["revision"] or ""
     cmd = [
         sys.executable,
@@ -52,6 +53,8 @@ def build_command() -> list[str]:
         cmd.append("--enable-auto-tool-choice")
     if updates["tool_call_parser"]:
         cmd.extend(["--tool-call-parser", str(updates["tool_call_parser"])])
+    if quantization:
+        cmd.extend(["--quantization", quantization])
     if revision:
         cmd.extend(["--revision", revision])
     if api_key: