[Refactor][Config] Remove weight NZ environment variable

docs/source/tutorials/models/DeepSeekOCR2.md

@@ -124,7 +124,6 @@ Run the following script to execute online inference.
 #!/bin/sh
 
 export VLLM_USE_V1=1
-export VLLM_ASCEND_ENABLE_NZ=0
 export TOKENIZERS_PARALLELISM=false
 export PYTORCH_NPU_ALLOC_CONF="expandable_segments:True"
 export TASK_QUEUE_ENABLE=1
@@ -140,6 +139,7 @@ vllm serve /root/.cache/DeepSeek-OCR-2 \
     --gpu-memory-utilization 0.8 \
     --allowed-local-media-path / \
     --additional-config '{
+      "weight_nz_mode": 0,
       "enable_cpu_binding": true,
       "multistream_overlap_shared_expert": true,
       "ascend_compilation_config": {"fuse_qknorm_rope": false}

docs/source/tutorials/models/GLM5.2.md

@@ -1293,7 +1293,6 @@ The 1M context scenarios are validated on Atlas 800 A3 only; the A2 series is no
 Recommended command:
 
 ```shell
-export VLLM_ASCEND_ENABLE_NZ=1
 export HCCL_OP_EXPANSION_MODE="AIV"
 export OMP_PROC_BIND=false
 export OMP_NUM_THREADS=20
@@ -1322,7 +1321,7 @@ vllm serve <MODEL_PATH> \
   --decode-context-parallel-size 16 \
   --cp-kv-cache-interleave-size 128 \
   --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY", "cudagraph_capture_sizes": [4, 16, 128]}' \
-  --additional-config '{"enable_flashcomm1": true, "enable_dsa_cp": true, "ascend_compilation_config": {"enable_npugraph_ex": true}, "multistream_overlap_shared_expert": true, "enable_sparse_li_c8": true, "enable_cpu_binding": true}' \
+  --additional-config '{"enable_flashcomm1": true, "enable_dsa_cp": true, "ascend_compilation_config": {"enable_npugraph_ex": true}, "multistream_overlap_shared_expert": true, "enable_sparse_li_c8": true, "enable_cpu_binding": true,"weight_nz_mode":1}' \
   --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp", "enforce_eager": true}' \
   --quantization ascend \
   --enable-expert-parallel \
@@ -1333,7 +1332,7 @@ Key Parameter Descriptions (in addition to [Single-Node Deployment](#5111-single
 
 **1M-specific environment variables:**
 
-- `VLLM_ASCEND_ENABLE_NZ=1`: Enables NZ format memory layout for the C8 quantized tensors, required for the 1M context deployment.
+- `additional_config.weight_nz_mode=1`: Enables NZ format memory layout for the C8 quantized tensors, required for the 1M context deployment.
 - `VLLM_WORKER_MULTIPROC_METHOD=spawn`: Uses the spawn start method for multi-process workers (required in this scenario).
 
 **1M-specific vllm serve parameters:**
@@ -1359,7 +1358,6 @@ export HCCL_IF_IP=$local_ip
 export GLOO_SOCKET_IFNAME=$nic_name
 export TP_SOCKET_IFNAME=$nic_name
 export HCCL_SOCKET_IFNAME=$nic_name
-export VLLM_ASCEND_ENABLE_NZ=1
 export HCCL_OP_EXPANSION_MODE="AIV"
 export OMP_PROC_BIND=false
 export OMP_NUM_THREADS=20
@@ -1392,7 +1390,7 @@ vllm serve <MODEL_PATH> \
   --decode-context-parallel-size 8 \
   --cp-kv-cache-interleave-size 128 \
   --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
-  --additional-config '{"enable_flashcomm1": true, "enable_dsa_cp": true, "ascend_compilation_config": {"enable_npugraph_ex": true}, "multistream_overlap_shared_expert": true,"enable_sparse_li_c8": true, "enable_cpu_binding": true}' \
+  --additional-config '{"enable_flashcomm1": true, "enable_dsa_cp": true, "ascend_compilation_config": {"enable_npugraph_ex": true}, "multistream_overlap_shared_expert": true,"enable_sparse_li_c8": true, "enable_cpu_binding": true,"weight_nz_mode":1}' \
   --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp", "enforce_eager": true}' \
   --quantization ascend \
   --enable-expert-parallel \
@@ -1425,7 +1423,6 @@ export GLOO_SOCKET_IFNAME=$nic_name
 export TP_SOCKET_IFNAME=$nic_name
 export HCCL_SOCKET_IFNAME=$nic_name
 
-export VLLM_ASCEND_ENABLE_NZ=1
 export HCCL_OP_EXPANSION_MODE="AIV"
 export OMP_PROC_BIND=false
 export OMP_NUM_THREADS=20
@@ -1459,7 +1456,7 @@ vllm serve <MODEL_PATH> \
   --decode-context-parallel-size 8 \
   --cp-kv-cache-interleave-size 128 \
   --enforce-eager \
-  --additional-config '{"enable_flashcomm1": true, "enable_dsa_cp": true, "ascend_compilation_config": {"enable_npugraph_ex": true}, "multistream_overlap_shared_expert": true,"enable_sparse_li_c8": true, "enable_cpu_binding": true, "recompute_scheduler_enable": true}' \
+  --additional-config '{"enable_flashcomm1": true, "enable_dsa_cp": true, "ascend_compilation_config": {"enable_npugraph_ex": true}, "multistream_overlap_shared_expert": true,"enable_sparse_li_c8": true, "enable_cpu_binding": true, "recompute_scheduler_enable": true,"weight_nz_mode":1}' \
   --speculative-config '{"num_speculative_tokens": 1, "method": "deepseek_mtp", "enforce_eager": true}' \
   --quantization ascend \
   --enable-expert-parallel \
@@ -1498,7 +1495,6 @@ export HCCL_IF_IP=$local_ip
 export GLOO_SOCKET_IFNAME=$nic_name
 export TP_SOCKET_IFNAME=$nic_name
 export HCCL_SOCKET_IFNAME=$nic_name
-export VLLM_ASCEND_ENABLE_NZ=1
 export HCCL_OP_EXPANSION_MODE="AIV"
 export OMP_PROC_BIND=false
 export OMP_NUM_THREADS=20
@@ -1532,7 +1528,7 @@ vllm serve <MODEL_PATH> \
   --decode-context-parallel-size 8 \
   --cp-kv-cache-interleave-size 128 \
   --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
-  --additional-config '{"ascend_compilation_config": {"enable_npugraph_ex": true},"multistream_overlap_shared_expert": true,"enable_sparse_li_c8": true, "enable_cpu_binding": true, "recompute_scheduler_enable": true}' \
+  --additional-config '{"ascend_compilation_config": {"enable_npugraph_ex": true},"multistream_overlap_shared_expert": true,"enable_sparse_li_c8": true, "enable_cpu_binding": true, "recompute_scheduler_enable": true,"weight_nz_mode":1}' \
   --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp", "enforce_eager": true}' \
   --quantization ascend \
   --enable-expert-parallel \

docs/source/user_guide/configuration/additional_config.md

@@ -85,7 +85,7 @@ The following table lists additional configuration options available in vLLM Asc
 | `enable_flashcomm1`                 | bool | `False` | Whether to enable FlashComm1 optimization. Can also be configured via the `VLLM_ASCEND_ENABLE_FLASHCOMM1` environment variable during the migration period. |
 | `msmonitor_use_daemon`              | bool | `False` | Whether to use daemon mode for msmonitor. Can also be configured via the `MSMONITOR_USE_DAEMON` environment variable during the migration period. |
 | `enable_mlapo`                      | bool | `True`  | Whether to enable MLAPO (Model Layer-wise Adaptive Parallel Optimization). Can also be configured via the `VLLM_ASCEND_ENABLE_MLAPO` environment variable during the migration period. |
-| `weight_nz_mode`                    | int  | `1`     | Weight NZ mode. Can also be configured via the `VLLM_ASCEND_ENABLE_NZ` environment variable during the migration period. |
+| `weight_nz_mode`                    | int  | `1`     | Weight NZ mode. `0` disables NZ, `1` enables NZ only for quantized weights, and `2` also enables NZ for BF16/FP16 weights when supported. The legacy `VLLM_ASCEND_ENABLE_NZ` environment variable is no longer supported. |
 | `enable_fused_mc2`                  | int  | `0`     | Fused MC2 configuration. Can also be configured via the `VLLM_ASCEND_ENABLE_FUSED_MC2` environment variable during the migration period. |
 | `enable_transpose_kv_cache_by_block`| bool | `True`  | Whether to enable transpose KV cache by block. Can also be configured via the `VLLM_ASCEND_FUSION_OP_TRANSPOSE_KV_CACHE_BY_BLOCK` environment variable during the migration period. |
 | `enable_dsa_cp`                     | bool | `False` | Whether to enable dsa_cp for DeepSeek V3.2, DeepSeek V4, and other models with the same architecture. This feature depends on FlashComm1. Please ensure that FlashComm1 is enabled before enabling this feature.|
@@ -244,9 +244,9 @@ ShortRequestFirst is a waiting-queue policy for FCFS synchronous or asynchronous
 
 **rl_config**
 
-`rl_config` is a one-click RL mode switch. When `enabled` is `true`, it refreshes the global Ascend configuration on every initialization, forces `AscendConfig.weight_nz_mode=0`, synchronizes `VLLM_ASCEND_ENABLE_NZ=0`, sets `VLLM_SERVER_DEV_MODE=1`, and removes the `expandable_segments` entry from `PYTORCH_NPU_ALLOC_CONF` with an informational log. These fixed RL behaviors are not configurable as `rl_config` sub-fields. When `enabled` is `false`, all other sub-fields are ignored.
+`rl_config` is a one-click RL mode switch. When `enabled` is `true`, it refreshes the global Ascend configuration on every initialization, forces `AscendConfig.weight_nz_mode=0`, sets `VLLM_SERVER_DEV_MODE=1`, and removes the `expandable_segments` entry from `PYTORCH_NPU_ALLOC_CONF` with an informational log. These fixed RL behaviors are not configurable as `rl_config` sub-fields. When `enabled` is `false`, all other sub-fields are ignored.
 
-When RL mode is enabled, its fixed NZ and developer-endpoint settings take precedence over top-level configuration and environment variables. `VLLM_BATCH_INVARIANT=1` remains enabled when `rl_config.enable_batch_invariant` is false.
+When RL mode is enabled, its fixed NZ setting takes precedence over the top-level `weight_nz_mode` configuration. `VLLM_BATCH_INVARIANT=1` remains enabled when `rl_config.enable_batch_invariant` is false.
 
 | Name | Type | Default | Description |
 | ---- | ---- | ------- | ----------- |

docs/source/user_guide/feature_guide/sleep_mode.md

@@ -100,7 +100,6 @@ The following is a simple example of how to use sleep mode.
 
     os.environ["VLLM_USE_MODELSCOPE"] = "True"
     os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
-    os.environ["VLLM_ASCEND_ENABLE_NZ"] = "0"
 
     if __name__ == "__main__":
         prompt = "How are you?"
@@ -109,7 +108,11 @@ The following is a simple example of how to use sleep mode.
         print(f"Free memory before sleep: {free / 1024 ** 3:.2f} GiB")
         # record npu memory use baseline in case other process is running
         used_bytes_baseline = total - free
-        llm = LLM("Qwen/Qwen2.5-0.5B-Instruct", enable_sleep_mode=True)
+        llm = LLM(
+            "Qwen/Qwen2.5-0.5B-Instruct",
+            enable_sleep_mode=True,
+            additional_config={"weight_nz_mode": 0},
+        )
         sampling_params = SamplingParams(temperature=0, max_tokens=10)
         output = llm.generate(prompt, sampling_params)
 
@@ -137,9 +140,10 @@ The following is a simple example of how to use sleep mode.
     export VLLM_SERVER_DEV_MODE="1"
     export VLLM_WORKER_MULTIPROC_METHOD="spawn"
     export VLLM_USE_MODELSCOPE="True"
-    export VLLM_ASCEND_ENABLE_NZ="0"
 
-    vllm serve Qwen/Qwen2.5-0.5B-Instruct --enable-sleep-mode
+    vllm serve Qwen/Qwen2.5-0.5B-Instruct \
+        --enable-sleep-mode \
+        --additional-config='{"weight_nz_mode": 0}'
 
     # after serving is up, post to these endpoints.
     # /sleep reads level from the query string (JSON body is ignored).

examples/rl/rlhf_async_new_apis.py

@@ -190,9 +190,6 @@ def generate(self, token_ids: list[int], max_new_tokens: int) -> list[int]:
     "VLLM_BATCH_INVARIANT": "1",
     "HCCL_DETERMINISTIC": "strict",
     "LCCL_DETERMINISTIC": "1",
-    # Disable FRACTAL_NZ mode (also handled by batch invariance override_envs)
-    "VLLM_ASCEND_ENABLE_NZ": "0",
-    "VLLM_ASCEND_ENABLE_MATMUL_ALLREDUCE": "0",
     # Enable expandable segments for PyTorch NPU allocator
     "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
 }
@@ -210,6 +207,7 @@ def generate(self, token_ids: list[int], max_new_tokens: int) -> list[int]:
     max_model_len=8192,
     distributed_executor_backend="ray",
     gpu_memory_utilization=0.75,
+    additional_config={"weight_nz_mode": 0},
     weight_transfer_config=WeightTransferConfig(backend="hccl"),
 )
 

tests/e2e/nightly/single_node/models/configs/DeepSeek-V3.2-W8A8-DCP.yaml

@@ -6,7 +6,6 @@ test_cases:
   - name: "DeepSeek-V3.2-W8A8-DCP-replicated-indexer"
     model: "vllm-ascend/DeepSeek-V3.2-W8A8"
     envs:
-      VLLM_ASCEND_ENABLE_NZ: "1"
       HCCL_OP_EXPANSION_MODE: "AIV"
       OMP_PROC_BIND: "false"
       OMP_NUM_THREADS: "20"
@@ -63,7 +62,7 @@ test_cases:
       - "--compilation-config"
       - '{"cudagraph_mode": "FULL_DECODE_ONLY", "cudagraph_capture_sizes":[4, 16, 64, 128]}'
       - "--additional-config"
-      - '{"enable_dsa_cp": true, "ascend_compilation_config":{"enable_npugraph_ex": true, "enable_static_kernel": false}, "multistream_overlap_shared_expert": true, "enable_mc2_hierarchy_comm": false, "enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true, "enable_cpu_binding": true, "recompute_scheduler_enable": false}'
+      - '{"enable_dsa_cp": true, "ascend_compilation_config":{"enable_npugraph_ex": true, "enable_static_kernel": false}, "multistream_overlap_shared_expert": true, "enable_mc2_hierarchy_comm": false, "enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true, "enable_cpu_binding": true, "recompute_scheduler_enable": false,"weight_nz_mode":1}'
       - "--speculative-config"
       - '{"num_speculative_tokens": 3, "method": "deepseek_mtp"}'
     test_content: []

tests/e2e/nightly/single_node/models/configs/Kimi-K2.5.yaml

@@ -12,7 +12,6 @@ _envs: &envs
   TASK_QUEUE_ENABLE: "1"
   VLLM_ASCEND_ENABLE_MLAPO: "1"
   VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
-  VLLM_ASCEND_ENABLE_NZ: "1"
 
 _server_cmd: &server_cmd
   - "--enable-expert-parallel"
@@ -44,7 +43,7 @@ _server_cmd: &server_cmd
   - "--speculative-config"
   - '{"method":"eagle3", "model":"lightseekorg/kimi-k2.5-eagle3", "num_speculative_tokens":3}'
   - "--additional-config"
-  - '{"enable_shared_expert_dp":true}'
+  - '{"enable_shared_expert_dp":true,"weight_nz_mode":1}'
   - "--mm-processor-cache-gb"
   - "0"
   - "--mm-encoder-tp-mode"

tests/e2e/nightly/single_node/models/configs/MiniMax-M2.5-w8a8-QuaRot-A2.yaml

@@ -13,7 +13,6 @@ test_cases:
       OMP_NUM_THREADS: "1"
       TASK_QUEUE_ENABLE: "1"
       VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
-      VLLM_ASCEND_ENABLE_NZ: "1"
       VLLM_TORCH_PROFILER_WITH_STACK: "0"
       VLLM_TORCH_PROFILER_DIR: "./profile"
       VLLM_USE_MODELSCOPE: "true"
@@ -52,7 +51,7 @@ test_cases:
       - "--compilation-config"
       - '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
       - "--additional-config"
-      - '{"enable_cpu_binding": true, "enable_npugraph_ex": true, "enable_static_kernel": true}'
+      - '{"enable_cpu_binding": true, "enable_npugraph_ex": true, "enable_static_kernel": true,"weight_nz_mode":1}'
     benchmarks:
       acc:
         case_type: accuracy

tests/e2e/nightly/single_node/models/configs/Qwen3-VL-235B-A22B-Instruct-W8A8.yaml

@@ -11,7 +11,6 @@ _envs: &envs
   PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
   VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
   VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
-  VLLM_ASCEND_ENABLE_NZ: "2"
   VLLM_ASCEND_BALANCE_SCHEDULING: "1"
   SERVER_PORT: "DEFAULT_PORT"
 
@@ -66,5 +65,7 @@ test_cases:
     server_cmd_extra:
       - "--compilation_config"
       - '{"cudagraph_mode": "FULL_DECODE_ONLY", "cudagraph_capture_sizes": [1,2,4,8,16,24,32]}'
+      - --additional-config
+      - '{"weight_nz_mode":2}'
     benchmarks:
       <<: *benchmarks

tests/e2e/nightly/single_node/models/configs/Qwen3.5-27B-w8a8-A2.yaml

@@ -13,7 +13,6 @@ test_cases:
       OMP_NUM_THREADS: "1"
       TASK_QUEUE_ENABLE: "1"
       VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE: "1"
-      VLLM_ASCEND_ENABLE_NZ: "1"
       VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
       SERVER_PORT: "DEFAULT_PORT"
     server_cmd:
@@ -33,7 +32,7 @@ test_cases:
       - "--gpu-memory-utilization"
       - "0.95"
       - "--additional-config"
-      - '{"enable_cpu_binding":true, "enable_weight_nz_layout":true}'
+      - '{"enable_cpu_binding":true, "enable_weight_nz_layout":true,"weight_nz_mode":1}'
       - "--speculative_config"
       - '{"method": "qwen3_5_mtp", "num_speculative_tokens": 3,"enforce_eager": true}'
       - "--compilation-config"

tests/e2e/pull_request/one_card/rlhf/conftest.py

@@ -93,7 +93,6 @@ def server(
     env = {
         **os.environ,
         "VLLM_SERVER_DEV_MODE": "1",
-        "VLLM_ASCEND_ENABLE_NZ": "0",
         "HF_HUB_OFFLINE": "1",
     }
     base = _DUMMY_ARGS if dummy_weights else _BASE_ARGS
@@ -107,6 +106,8 @@ def server(
         str(port),
         "--served-model-name",
         "m",
+        "--additional-config",
+        '{"weight_nz_mode": 0}',
         *(base + (extra_args or [])),
     ]
     # Establish the test process's NPU device context while the card is still

tests/e2e/pull_request/one_card/rlhf/state_transitions/test_sleep_wake.py

@@ -28,7 +28,7 @@
   --enable-sleep-mode   KV cache allocated via CuMemAllocator; without this
                         flag sleep/wake are no-ops and the bug cannot trigger.
   VLLM_SERVER_DEV_MODE=1
-  VLLM_ASCEND_ENABLE_NZ=0
+  --additional-config '{"weight_nz_mode": 0}'
 """
 
 import requests

tests/e2e/pull_request/one_card/test_npu_ipc_weight_transfer.py

@@ -97,6 +97,8 @@ def test_npu_ipc_weight_transfer_updates_server_weights():
         "--port",
         str(port),
         "--trust-remote-code",
+        "--additional-config",
+        '{"weight_nz_mode": 0}',
     ]
     # VLLM_SERVER_DEV_MODE registers the dev endpoints; insecure serialization
     # lets the server unpickle the IPC handles sent over HTTP. Pin the server to
@@ -105,7 +107,6 @@ def test_npu_ipc_weight_transfer_updates_server_weights():
         "VLLM_SERVER_DEV_MODE": "1",
         "VLLM_ALLOW_INSECURE_SERIALIZATION": "1",
         "ASCEND_RT_VISIBLE_DEVICES": str(INFERENCE_DEVICE_INDEX),
-        "VLLM_ASCEND_ENABLE_NZ": "0",
     }
 
     with RemoteOpenAIServer(

tests/e2e/pull_request/one_card/test_xlite.py

@@ -23,15 +23,11 @@
 
 # ruff: noqa: E501
 
-import os
-
 import pytest
 
 from tests.e2e.conftest import wait_until_npu_memory_free
 from tests.e2e.pull_request.utils import PROMPTS_SHORT, compare_logprobs
 
-os.environ["VLLM_ASCEND_ENABLE_NZ"] = "2"
-
 MODELS: list[str] = ["Qwen/Qwen3-0.6B"]
 
 
@@ -52,7 +48,10 @@ def test_models_with_xlite_decode_only(model: str):
         "model_name": model,
         "max_model_len": 1024,
         "block_size": 128,
-        "additional_config": {"xlite_graph_config": {"enabled": True, "full_mode": False}},
+        "additional_config": {
+            "weight_nz_mode": 2,
+            "xlite_graph_config": {"enabled": True, "full_mode": False},
+        },
     }
     compare_logprobs(runner_kwargs=runner_kwargs, prompts=PROMPTS_SHORT)
 
@@ -74,6 +73,9 @@ def test_models_with_xlite_full_mode(model: str):
         "model_name": model,
         "max_model_len": 1024,
         "block_size": 128,
-        "additional_config": {"xlite_graph_config": {"enabled": True, "full_mode": True}},
+        "additional_config": {
+            "weight_nz_mode": 2,
+            "xlite_graph_config": {"enabled": True, "full_mode": True},
+        },
     }
     compare_logprobs(runner_kwargs=runner_kwargs, prompts=PROMPTS_SHORT)

tests/e2e/pull_request/two_card/test_hccl_weight_transfer.py

@@ -188,14 +188,15 @@ def test_hccl_weight_transfer_updates_server_weights():
         "--port",
         str(port),
         "--trust-remote-code",
+        "--additional-config",
+        '{"weight_nz_mode": 0}',
     ]
     # The dev-mode endpoints (/init_weight_transfer_engine, /update_weights,
     # /pause, /resume, ...) are only registered when VLLM_SERVER_DEV_MODE=1.
     # Pin the server to NPU 0 so the trainer can own NPU 1 exclusively.
     env_dict = {
         "VLLM_SERVER_DEV_MODE": "1",
         "ASCEND_RT_VISIBLE_DEVICES": "0",
-        "VLLM_ASCEND_ENABLE_NZ": "0",
     }
 
     _log(f"starting server on port {port} (device 0, dummy weights) ...")

tests/e2e/pull_request/two_card/test_xlite.py

@@ -28,8 +28,6 @@
 from tests.e2e.conftest import DPVllmRunner, VllmRunner, wait_until_npu_memory_free
 from tests.e2e.pull_request.utils import PROMPTS_SHORT
 
-os.environ["VLLM_ASCEND_ENABLE_NZ"] = "2"
-
 MODELS: list[str] = ["Qwen/Qwen3-30B-A3B"]
 TPDP_SIZES: list[tuple[int, int]] = [(2, 1), (1, 2)]
 
@@ -58,7 +56,10 @@ def test_models_with_xlite_decode_only(model: str, tpdp: tuple[int, int]):
         enable_expert_parallel=True,
         block_size=128,
         max_model_len=2048,
-        additional_config={"xlite_graph_config": {"enabled": True, "full_mode": False}},
+        additional_config={
+            "weight_nz_mode": 2,
+            "xlite_graph_config": {"enabled": True, "full_mode": False},
+        },
     ) as vllm_model:
         outputs = vllm_model.generate_greedy(PROMPTS_SHORT, 3)
 
@@ -89,7 +90,10 @@ def test_models_with_xlite_full_mode(model: str, tpdp: tuple[int, int]):
         enable_expert_parallel=True,
         block_size=128,
         max_model_len=2048,
-        additional_config={"xlite_graph_config": {"enabled": True, "full_mode": True}},
+        additional_config={
+            "weight_nz_mode": 2,
+            "xlite_graph_config": {"enabled": True, "full_mode": True},
+        },
     ) as vllm_model:
         outputs = vllm_model.generate_greedy(PROMPTS_SHORT, 3)
 

tests/e2e/weekly/single_node/models/test_qwen3_30b_acc.py

@@ -85,7 +85,6 @@ async def test_models(model: str, tp_size: int) -> None:
         "HCCL_BUFFSIZE": "1024",
         "OMP_NUM_THREADS": "1",
         "PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
-        "VLLM_ASCEND_ENABLE_NZ": "2",
         "MOONCAKE_CONFIG_PATH": "mooncake.json",
     }
     if tp_size != 1: