.

Author: ZiyiTsang

areal/engine/sglang_remote.py

@@ -88,11 +88,13 @@ def parse_generation_response(
         # Extract routed_experts information if available
         routed_experts = meta_info.get("routed_experts", None)
         if routed_experts is not None:
-            num_sgl_token=meta_info['prompt_tokens'] + meta_info['completion_tokens']-1
+            num_sgl_token = (
+                meta_info["prompt_tokens"] + meta_info["completion_tokens"] - 1
+            )
             # Extract expert_id and reshape to (num_sgl_token, num_layers*expert_top_k)
             routed_experts = np.frombuffer(
                 pybase64.b64decode(routed_experts.encode("utf-8")), dtype=np.int32
-            ).reshape(num_sgl_token,-1)
+            ).reshape(num_sgl_token, -1)
 
         if stop_reason == "abort" and stop_message.startswith("Abort before prefill"):
             return HttpGenerationResult(

areal/infra/launcher/sglang_server.py

@@ -12,7 +12,6 @@
 from areal.api.alloc_mode import AllocationMode
 from areal.api.cli_args import (
     ClusterSpecConfig,
-    GenerationHyperparameters,
     NameResolveConfig,
     SGLangConfig,
     parse_cli_args,

docs/cli_reference.md

@@ -575,6 +575,7 @@ https://github.com/sgl-project/sglang for detailed documentation.
 | `enable_metrics`                  | boolean                 | `True`       | -           |
 | `decode_log_interval`             | integer                 | `1`          | -           |
 | `enable_multithread_load`         | boolean                 | `False`      | -           |
+| `enable_return_routed_experts`    | boolean                 | `False`      | -           |
 
 (section-v-llm)=
 
@@ -747,7 +748,7 @@ Configuration for SwanLab experiment tracking and monitoring.
 | `config`  | `dict` \| None | `None`                    | -           |
 | `logdir`  | string \| None | `None`                    | -           |
 | `mode`    | string \| None | `"disabled"`              | -           |
-| `api_key` | string \| None | `None` | -           |
+| `api_key` | string \| None | `"4KHJ20cRg6fuVkvwvGaXD"` | -           |
 
 (section-tensor-board)=