Simplify tests & Fix trtllm custom allreduce registration (#4252)

python/sglang/srt/_custom_ops.py

@@ -6,10 +6,12 @@ from typing import List, Tuple
 import torch
 import torch.library
 
-from sglang.srt.utils import is_hip, is_hpu
+from sglang.srt.utils import get_bool_env_var, is_hip, is_hpu
 
 logger = logging.getLogger(__name__)
-use_vllm_custom_allreduce = os.environ.get("USE_VLLM_CUSTOM_ALLREDUCE", default=True)
+use_vllm_custom_allreduce = get_bool_env_var(
+    "USE_VLLM_CUSTOM_ALLREDUCE", default="true"
+)
 
 if not is_hpu():
     # ROCm does not use vllm custom allreduce

python/sglang/srt/layers/attention/flashinfer_backend.py

@@ -22,7 +22,7 @@ from sglang.srt.layers.attention.utils import create_flashinfer_kv_indices_trito
 from sglang.srt.layers.dp_attention import get_attention_tp_size
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
 from sglang.srt.speculative.eagle_utils import EagleDraftInput, EagleVerifyInput
-from sglang.srt.utils import is_flashinfer_available
+from sglang.srt.utils import get_bool_env_var, is_flashinfer_available
 
 if TYPE_CHECKING:
     from sglang.srt.layers.radix_attention import RadixAttention

python/sglang/srt/model_loader/loader.py

@@ -48,6 +48,7 @@ from sglang.srt.model_loader.weight_utils import (
     safetensors_weights_iterator,
 )
 from sglang.srt.utils import (
+    get_bool_env_var,
     get_device_capability,
     is_pin_memory_available,
     set_weight_attrs,
@@ -197,7 +198,7 @@ class DefaultModelLoader(BaseModelLoader):
 
         Returns the path to the downloaded model, or None if the model is not
         downloaded from ModelScope."""
-        if os.environ.get("SGLANG_USE_MODELSCOPE", None) == "True":
+        if get_bool_env_var("SGLANG_USE_MODELSCOPE"):
             # download model from ModelScope hub,
             # lazy import so that modelscope is not required for normal use.
             # pylint: disable=C.