fix: support sglang>=0.5.4

draft/qwen2.py

@@ -43,7 +43,14 @@ from sglang.srt.layers.vocab_parallel_embedding import (
     ParallelLMHead,
     VocabParallelEmbedding,
 )
-from sglang.srt.managers.schedule_batch import global_server_args_dict
+try:
+    from sglang.srt.managers.schedule_batch import global_server_args_dict
+except ImportError:
+    global_server_args_dict = None
+try:
+    from sglang.srt.server_args import get_global_server_args
+except ImportError:
+    get_global_server_args = None
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import (
     default_weight_loader,
@@ -273,7 +280,11 @@ class Qwen2Model(nn.Module):
                 config.vocab_size,
                 config.hidden_size,
                 quant_config=quant_config,
-                enable_tp=not global_server_args_dict["enable_dp_attention"],
+                enable_tp=not (
+                    global_server_args_dict["enable_dp_attention"]
+                    if global_server_args_dict is not None
+                    else get_global_server_args().enable_dp_attention
+                ),
                 prefix=add_prefix("embed_tokens", prefix),
             )
         else: