update

Signed-off-by: shen-shanshan <467638484@qq.com>

Author: shen-shanshan

vllm_ascend/ascend_forward_context.py

@@ -8,7 +8,7 @@
 from vllm.distributed import get_dp_group, get_ep_group, get_tp_group
 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.platform import NPUPlatform
 
 
@@ -26,15 +26,15 @@ 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/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/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/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)

vllm_ascend/quantization/w8a8_dynamic.py

@@ -23,7 +23,7 @@
 from vllm.distributed import GroupCoordinator, get_ep_group
 from vllm.forward_context import get_forward_context
 
-import vllm_ascend.envs as envs
+import vllm_ascend.envs as envs_ascend
 from vllm_ascend.ascend_config import get_ascend_config
 from vllm_ascend.ascend_forward_context import FusedMoEState
 from vllm_ascend.distributed.parallel_state import get_mc2_group
@@ -1019,7 +1019,7 @@ def process_weights_after_loading(self, layer):
                 1, 2).contiguous()
             layer.w2_weight.data = layer.w2_weight.data.transpose(
                 1, 2).contiguous()
-        if envs.VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP:
+        if envs_ascend.VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP:
             torch_npu.npu_format_cast_(layer.w2_weight, ACL_FORMAT_FRACTAL_NZ)
         layer.w13_weight_scale.data = layer.w13_weight_scale.data.view(
             layer.w13_weight_scale.data.shape[0], -1)

vllm_ascend/utils.py

@@ -31,7 +31,7 @@
 from torch_npu.npu.streams import Event
 from vllm.logger import logger
 
-import vllm_ascend.envs as envs
+import vllm_ascend.envs as envs_ascend
 from vllm_ascend.ascend_config import get_ascend_config
 
 if TYPE_CHECKING:
@@ -232,7 +232,7 @@ def find_hccl_library() -> str:
     After importing `torch`, `libhccl.so` can be
     found by `ctypes` automatically.
     """
-    so_file = envs.HCCL_SO_PATH
+    so_file = envs_ascend.HCCL_SO_PATH
 
     # manually load the hccl library
     if so_file:
@@ -273,8 +273,8 @@ def adapt_patch(is_global_patch: bool = False):
 
 @functools.cache
 def vllm_version_is(target_vllm_version: str):
-    if envs.VLLM_VERSION is not None:
-        vllm_version = envs.VLLM_VERSION
+    if envs_ascend.VLLM_VERSION is not None:
+        vllm_version = envs_ascend.VLLM_VERSION
     else:
         import vllm
         vllm_version = vllm.__version__
@@ -385,7 +385,7 @@ def destroy(self):
 
     @contextmanager
     def capture_async(self, duration_tag: str):
-        if not envs.VLLM_ASCEND_MODEL_EXECUTE_TIME_OBSERVE:
+        if not envs_ascend.VLLM_ASCEND_MODEL_EXECUTE_TIME_OBSERVE:
             yield
             return
 
@@ -403,7 +403,7 @@ def capture_async(self, duration_tag: str):
     def pop_captured_sync(self) -> dict:
         """Pop and synchronize all events in the observation list"""
         durations: dict[str, float] = {}
-        if not envs.VLLM_ASCEND_MODEL_EXECUTE_TIME_OBSERVE:
+        if not envs_ascend.VLLM_ASCEND_MODEL_EXECUTE_TIME_OBSERVE:
             return durations
 
         while self._observations:
@@ -437,7 +437,7 @@ def get_rm_router_logits_state(ep_size: int, dp_size: int,
     # the fusion operator torch_npu.npu_grouped_matmul_finalize_routing called by allgather ep
     # only supports deepseek v3/r1
     if dp_size > 1:
-        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 True
         elif ep_size == 1 and is_deepseek_v3_r1:
@@ -451,7 +451,7 @@ def get_rm_router_logits_state(ep_size: int, dp_size: int,
 def get_all_reduce_merge_state(ep_size: int, 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 True
     elif ep_size == 1 and is_deepseek_v3_r1:

vllm_ascend/worker/worker_v1.py

@@ -24,7 +24,7 @@
 import torch.nn as nn
 import torch_npu
 from torch_npu.op_plugin.atb._atb_ops import _register_atb_extensions
-from vllm import envs
+from vllm import envs_vllm
 from vllm.config import VllmConfig
 from vllm.distributed import (ensure_model_parallel_initialized,
                               init_distributed_environment)
@@ -317,8 +317,8 @@ def _init_worker_distributed_environment(self) -> None:
     def _init_profiler(self):
         # Torch profiler. Enabled and configured through env vars:
         # VLLM_TORCH_PROFILER_DIR=/path/to/save/trace
-        if envs.VLLM_TORCH_PROFILER_DIR:
-            torch_profiler_trace_dir = envs.VLLM_TORCH_PROFILER_DIR
+        if envs_vllm.VLLM_TORCH_PROFILER_DIR:
+            torch_profiler_trace_dir = envs_vllm.VLLM_TORCH_PROFILER_DIR
             logger.info("Profiling enabled. Traces will be saved to: %s",
                         torch_profiler_trace_dir)