[Misc] Remove redundant imported envs, using envs_ascend instead (#2193)

### What this PR does / why we need it?
Remove redundant imported `envs`, using `envs_ascend` instead.

```python
import vllm.envs as envs_vllm
import vllm_ascend.envs as envs_ascend
```

- vLLM version: v0.10.0
- vLLM main:
vllm-project/vllm@71683ca

---------

Signed-off-by: shen-shanshan <[email protected]>

Author: shen-shanshan

tests/ut/patch/worker/patch_common/test_patch_linear.py

@@ -5,8 +5,8 @@
 import vllm
 from pytest_mock import MockerFixture
 
+import vllm_ascend.envs as envs_ascend
 from tests.ut.base import PytestBase
-from vllm_ascend import envs
 from vllm_ascend.patch.worker.patch_common import patch_linear
 
 
@@ -158,10 +158,10 @@ def test_calc_output(
         assert torch.allclose(ret, expected)
 
     def test_enable_allreduce_matmul(self, mocker: MockerFixture):
-        mocker.patch.object(envs,
+        mocker.patch.object(envs_ascend,
                             "VLLM_ASCEND_ENABLE_MATMUL_ALLREDUCE",
                             new=True)
         reload(patch_linear)
-        assert envs.VLLM_ASCEND_ENABLE_MATMUL_ALLREDUCE
+        assert envs_ascend.VLLM_ASCEND_ENABLE_MATMUL_ALLREDUCE
         assert id(vllm.model_executor.layers.linear.RowParallelLinear) == id(
             patch_linear.AscendRowParallelLinear)

tests/ut/test_envs.py

@@ -15,25 +15,26 @@
 import inspect
 import os
 
+import vllm_ascend.envs as envs_ascend
 from tests.ut.base import TestBase
-from vllm_ascend import envs
 
 
 class TestEnvVariables(TestBase):
 
     def setUp(self):
-        self.env_vars = list(envs.env_variables.keys())
+        self.env_vars = list(envs_ascend.env_variables.keys())
 
     def test_env_vars_behavior(self):
         for var_name in self.env_vars:
             with self.subTest(var=var_name):
                 original_val = os.environ.get(var_name)
-                var_handler = envs.env_variables[var_name]
+                var_handler = envs_ascend.env_variables[var_name]
 
                 try:
                     if var_name in os.environ:
                         del os.environ[var_name]
-                    self.assertEqual(getattr(envs, var_name), var_handler())
+                    self.assertEqual(getattr(envs_ascend, var_name),
+                                     var_handler())
 
                     handler_source = inspect.getsource(var_handler)
                     if 'int(' in handler_source:
@@ -45,7 +46,7 @@ def test_env_vars_behavior(self):
 
                     for test_val in test_vals:
                         os.environ[var_name] = test_val
-                        self.assertEqual(getattr(envs, var_name),
+                        self.assertEqual(getattr(envs_ascend, var_name),
                                          var_handler())
 
                 finally:
@@ -55,7 +56,7 @@ def test_env_vars_behavior(self):
                         os.environ[var_name] = original_val
 
     def test_dir_and_getattr(self):
-        self.assertEqual(sorted(envs.__dir__()), sorted(self.env_vars))
+        self.assertEqual(sorted(envs_ascend.__dir__()), sorted(self.env_vars))
         for var_name in self.env_vars:
             with self.subTest(var=var_name):
-                getattr(envs, var_name)
+                getattr(envs_ascend, var_name)

vllm_ascend/ascend_forward_context.py

@@ -9,7 +9,7 @@
                               get_tensor_model_parallel_world_size)
 from vllm.forward_context import get_forward_context, set_forward_context
 
-import vllm_ascend.envs as envs
+import vllm_ascend.envs as envs_ascend
 from vllm_ascend.distributed.moe_comm_method import MoECommMethod
 
 
@@ -27,15 +27,15 @@ def _get_fused_moe_state(ep_size: int, with_prefill: bool,
                          is_deepseek_v3_r1: bool):
     # the fusion operator torch_npu.npu_grouped_matmul_finalize_routing called by allgather ep
     # only supports deepseek v3/r1
-    if (envs.VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP and ep_size > 1
+    if (envs_ascend.VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP and ep_size > 1
             and is_deepseek_v3_r1):
         return FusedMoEState.AllGatherEP
     elif ep_size == 1:
         if with_prefill:
             return FusedMoEState.NaiveMulticast
         else:
             return FusedMoEState.AllGather
-    elif envs.VLLM_ASCEND_ENABLE_MOE_ALL2ALL_SEQ:
+    elif envs_ascend.VLLM_ASCEND_ENABLE_MOE_ALL2ALL_SEQ:
         # MC2 Dispatch/Combine performs better than alltoall_seq in decoding stage.
         return (FusedMoEState.All2AllSeq if
                 (ep_size < 16 or with_prefill) else FusedMoEState.MC2)

vllm_ascend/attention/mla_v1.py

@@ -14,7 +14,7 @@
                                                UnquantizedLinearMethod)
 from vllm.utils import cdiv, round_down
 
-from vllm_ascend import envs
+import vllm_ascend.envs as envs_ascend
 from vllm_ascend.ascend_config import get_ascend_config
 from vllm_ascend.attention.attention_v1 import AscendAttentionState
 from vllm_ascend.multistream.base import MSAttentionMetadataSplitConfig
@@ -1054,7 +1054,7 @@ def _forward_decode(
             # be removed after the torch_npu contains `torch_npu.atb.npu_multi_head_latent_attention` become
             # public available
             assert len(kv_c_and_k_pe_cache) > 1
-            if envs.VLLM_ASCEND_MLA_PA:
+            if envs_ascend.VLLM_ASCEND_MLA_PA:
                 attn_output = torch_npu.atb.npu_multi_head_latent_attention(
                     q_nope, q_pe, kv_c_and_k_pe_cache[0],
                     kv_c_and_k_pe_cache[1], attn_metadata.decode.block_table,

vllm_ascend/compilation/piecewise_backend.py

@@ -23,7 +23,7 @@
 
 import torch
 import torch.fx as fx
-import vllm.envs as envs
+import vllm.envs as envs_vllm
 from vllm.compilation.backends import VllmBackend
 from vllm.compilation.counter import compilation_counter
 from vllm.compilation.monitor import end_monitoring_torch_compile
@@ -93,7 +93,7 @@ def __init__(self, graph: fx.GraphModule, vllm_config: VllmConfig,
 
         self.sym_shape_indices = sym_shape_indices
 
-        self.is_debugging_mode = envs.VLLM_LOGGING_LEVEL == "DEBUG"
+        self.is_debugging_mode = envs_vllm.VLLM_LOGGING_LEVEL == "DEBUG"
 
         # the entries for different shapes that we need to either
         # compile or capture aclgraph

vllm_ascend/distributed/llmdatadist_c_mgr_connector.py

@@ -27,7 +27,7 @@
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.request import Request, RequestStatus
 
-from vllm_ascend import envs
+import vllm_ascend.envs as envs_ascend
 from vllm_ascend.utils import AscendSocVersion, get_ascend_soc_version
 
 TORCH_DTYPE_TO_NPU_DTYPE = {
@@ -181,7 +181,7 @@ def __init__(self, vllm_config: VllmConfig, engine_id: Optional[str]):
         dp_rank_local = self.vllm_config.parallel_config.data_parallel_rank_local
         tp_size = self.vllm_config.parallel_config.tensor_parallel_size
 
-        self.port = dp_rank_local * tp_size + envs.VLLM_LLMDD_RPC_PORT if dp_rank_local is not None else tp_size + envs.VLLM_LLMDD_RPC_PORT
+        self.port = dp_rank_local * tp_size + envs_ascend.VLLM_LLMDD_RPC_PORT if dp_rank_local is not None else tp_size + envs_ascend.VLLM_LLMDD_RPC_PORT
 
         self._reqs_need_recv: dict[str, tuple[Request, list[int]]] = {}
 
@@ -344,7 +344,7 @@ def __init__(self, vllm_config: VllmConfig):
 
     def listen_for_agent_metadata_req(self, event: threading.Event):
         assert self.local_agent_metadata is not None
-        port = envs.VLLM_LLMDD_RPC_PORT + self.local_dp_rank * self.tp_size + self.tp_rank if self.local_dp_rank is not None else envs.VLLM_LLMDD_RPC_PORT + self.tp_size + self.tp_rank
+        port = envs_ascend.VLLM_LLMDD_RPC_PORT + self.local_dp_rank * self.tp_size + self.tp_rank if self.local_dp_rank is not None else envs_ascend.VLLM_LLMDD_RPC_PORT + self.tp_size + self.tp_rank
         url = f"tcp://0.0.0.0:{port}"
         msg_encoder = msgspec.msgpack.Encoder()
         msg_decoder = msgspec.msgpack.Decoder()
@@ -427,9 +427,9 @@ def init_llm_datadist(self):
 
     def read_offline_rank_table(self):
         assert (
-            envs.DISAGGREGATED_PREFILL_RANK_TABLE_PATH
+            envs_ascend.DISAGGREGATED_PREFILL_RANK_TABLE_PATH
         ), "Please set path of rank_table to env variable DISAGGREGATED_PREFILL_RANK_TABLE_PATH"
-        rank_table_path = envs.DISAGGREGATED_PREFILL_RANK_TABLE_PATH
+        rank_table_path = envs_ascend.DISAGGREGATED_PREFILL_RANK_TABLE_PATH
         with open(rank_table_path, "r", encoding="utf-8") as f:
             global_rank_table = json.load(f)
         decode_device_list = global_rank_table["decode_device_list"]

vllm_ascend/models/__init__.py

@@ -1,6 +1,6 @@
 from vllm import ModelRegistry
 
-import vllm_ascend.envs as envs
+import vllm_ascend.envs as envs_ascend
 
 
 def register_model():
@@ -21,7 +21,7 @@ def register_model():
         "Qwen2VLForConditionalGeneration",
         "vllm_ascend.models.qwen2_vl:AscendQwen2VLForConditionalGeneration")
 
-    if envs.USE_OPTIMIZED_MODEL:
+    if envs_ascend.USE_OPTIMIZED_MODEL:
         ModelRegistry.register_model(
             "Qwen2_5_VLForConditionalGeneration",
             "vllm_ascend.models.qwen2_5_vl:AscendQwen2_5_VLForConditionalGeneration"
@@ -32,7 +32,7 @@ def register_model():
             "vllm_ascend.models.qwen2_5_vl_without_padding:AscendQwen2_5_VLForConditionalGeneration_Without_Padding"
         )
 
-    if envs.VLLM_ASCEND_ENABLE_DBO:
+    if envs_ascend.VLLM_ASCEND_ENABLE_DBO:
         ModelRegistry.register_model(
             "DeepseekV2ForCausalLM",
             "vllm_ascend.models.deepseek_dbo:CustomDeepseekDBOForCausalLM")

vllm_ascend/patch/platform/patch_common/patch_distributed.py

@@ -18,7 +18,7 @@
 # This file is a part of the vllm-ascend project.
 
 import torch
-import vllm.envs as envs
+import vllm.envs as envs_vllm
 from vllm.config import ParallelConfig
 
 from vllm_ascend.utils import is_310p
@@ -37,7 +37,7 @@ def parallel_config_get_dp_port(self) -> int:
     self.data_parallel_master_port += 1
 
     # NOTE: Get port from envs directly when using torchrun
-    port = envs.VLLM_DP_MASTER_PORT if envs.VLLM_DP_MASTER_PORT else answer
+    port = envs_vllm.VLLM_DP_MASTER_PORT if envs_vllm.VLLM_DP_MASTER_PORT else answer
     return port
 
 

vllm_ascend/patch/worker/patch_common/patch_linear.py

@@ -28,7 +28,7 @@
 from vllm.logger import logger
 from vllm.model_executor.layers.linear import RowParallelLinear
 
-from vllm_ascend import envs
+import vllm_ascend.envs as envs_ascend
 
 _HCOMM_INFO = None
 
@@ -142,6 +142,6 @@ def calc_output(self, input_parallel: torch.Tensor) -> torch.Tensor:
         return output
 
 
-if envs.VLLM_ASCEND_ENABLE_MATMUL_ALLREDUCE:
+if envs_ascend.VLLM_ASCEND_ENABLE_MATMUL_ALLREDUCE:
     logger.info("AscendRowParallelLinear: Matmul all-reduce is enabled. ")
     vllm.model_executor.layers.linear.RowParallelLinear = AscendRowParallelLinear

vllm_ascend/platform.py

@@ -20,7 +20,7 @@
 from typing import TYPE_CHECKING, Optional, Tuple
 
 import torch
-import vllm.envs as envs
+import vllm.envs as envs_vllm
 from torch.distributed import ProcessGroup
 from torch.distributed.distributed_c10d import PrefixStore
 from vllm.logger import logger
@@ -116,7 +116,7 @@ def clear_npu_memory(cls):
 
     @classmethod
     def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
-        if not envs.VLLM_USE_V1:
+        if not envs_vllm.VLLM_USE_V1:
             raise ValueError("vLLM Ascend does not support V0 engine.")
         # initialize ascend config from vllm additional_config
         ascend_config = init_ascend_config(vllm_config)