Cleanup: relocate temp_set_env and consolidate multi-device/CUDA helpers in common.py (#30180)

python/sglang/srt/distributed/parallel_state.py

@@ -2129,11 +2129,6 @@ def initialize_model_parallel(
             logger.info(
                 f"DCP enabled, dcp_size={decode_context_parallel_size}, tp_size={tensor_model_parallel_size}"
             )
-    else:
-        if get_tensor_model_parallel_rank() == 0:
-            logger.info(
-                f"DCP disabled, dcp_size={decode_context_parallel_size}, tp_size={tensor_model_parallel_size}"
-            )
 
     attn_dp_size = attention_data_parallel_size
     attn_cp_size = attention_context_model_parallel_size

python/sglang/srt/environ.py

@@ -24,35 +24,6 @@ def _default_hip() -> bool:
         return False
 
 
-@contextmanager
-def temp_set_env(*, allow_sglang: bool = False, **env_vars: Any):
-    """Temporarily set environment variables, restoring originals on exit.
-
-    By default, SGLANG_*/SGL_* keys are rejected — use ``Envs`` descriptors
-    for those.  Pass ``allow_sglang=True`` only for special env vars that
-    intentionally bypass ``environ.py``.
-    """
-    if not allow_sglang:
-        for key in env_vars:
-            if key.startswith("SGLANG_") or key.startswith("SGL_"):
-                raise ValueError("temp_set_env should not be used for sglang env vars")
-
-    backup = {key: os.environ.get(key) for key in env_vars}
-    try:
-        for key, value in env_vars.items():
-            if value is None:
-                os.environ.pop(key, None)
-            else:
-                os.environ[key] = str(value)
-        yield
-    finally:
-        for key, value in backup.items():
-            if value is None:
-                os.environ.pop(key, None)
-            else:
-                os.environ[key] = value
-
-
 class EnvField:
     _allow_set_name = True
 

python/sglang/srt/model_loader/loader.py

@@ -91,7 +91,7 @@ from sglang.srt.utils.common import is_cuda_alike
 DEFAULT_GPU_MEMORY_FRACTION_FOR_CALIBRATION = (
     0.8  # Reserve 20% GPU memory headroom for ModelOpt calibration
 )
-from sglang.srt.environ import envs, temp_set_env
+from sglang.srt.environ import envs
 from sglang.srt.model_loader.weight_utils import (
     buffered_multi_thread_safetensors_weights_iterator,
     download_safetensors_index_file_from_hf,
@@ -119,6 +119,7 @@ from sglang.srt.utils import (
     rank0_log,
     set_weight_attrs,
 )
+from sglang.srt.utils.common import temp_set_env
 
 if TYPE_CHECKING:
     from sglang.srt.configs.device_config import DeviceConfig

python/sglang/test/kits/mmmu_vlm_kit.py

@@ -7,8 +7,8 @@ import tempfile
 from pathlib import Path
 from types import SimpleNamespace
 
-from sglang.srt.environ import temp_set_env
 from sglang.srt.utils import kill_process_tree
+from sglang.srt.utils.common import temp_set_env
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,

test/registered/debug_utils/test_dumper.py

@@ -39,8 +39,8 @@ from sglang.srt.debug_utils.dumper import (
     get_tensor_info,
     get_truncated_value,
 )
-from sglang.srt.environ import temp_set_env
 from sglang.srt.utils import kill_process_tree
+from sglang.srt.utils.common import temp_set_env
 from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
 from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,

test/registered/model_loading/test_runai_model_loader.py

@@ -1,7 +1,7 @@
 import unittest
 
 import sglang as sgl
-from sglang.srt.environ import temp_set_env
+from sglang.srt.utils.common import temp_set_env
 from sglang.test.ci.ci_register import register_cuda_ci
 from sglang.test.test_utils import CustomTestCase