Support MiMo V2.5 with zigzag context parallelism (#29972)

2026-07-19 02:36:40 -07:00
parent 377c93d54e
commit 7a03d30149
17 changed files with 368 additions and 118 deletions
@@ -2095,7 +2095,13 @@ def _execute_server_warmup(server_args: ServerArgs):
    model_info = res.json()

    # Construct a warmup request (MLX: text warmup for VLM-advertising models; TODO: enable image warmup).
-    is_vlm = bool(model_info.get("has_image_understanding", False)) and not is_mps()
+    # A language-only worker may advertise VLM capability for encoder
+    # disaggregation, but its local warmup must stay on the text path.
+    is_vlm = (
+        bool(model_info.get("has_image_understanding", False))
+        and not server_args.language_only
+        and not is_mps()
+    )
    if model_info["is_generation"]:
        if is_vlm and not server_args.skip_tokenizer_init:
            request_name = "/v1/chat/completions"