[Refactor][Config] Remove MLAPO environment variable

docs/source/tutorials/models/DeepSeek-V3.1.md

@@ -183,8 +183,6 @@ Single-node deployment completes both Prefill and Decode within the same node. T
     export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
     export TASK_QUEUE_ENABLE=1
 
-    export VLLM_ASCEND_ENABLE_MLAPO=1
-
     vllm serve /weight/dsk-v3.1-w4a4_mlp-w8a8c8_attn-0618-full \
     --host 0.0.0.0 \
     --port 8015 \
@@ -202,7 +200,7 @@ Single-node deployment completes both Prefill and Decode within the same node. T
     --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}' \
     --speculative-config '{"num_speculative_tokens": 3,"method": "deepseek_mtp"}' \
     --quantization ascend \
-    --additional_config '{"enable_cpu_binding": true, "multistream_overlap_shared_expert": true, "enable_balance_scheduling": true}' \
+    --additional_config '{"enable_cpu_binding": true, "multistream_overlap_shared_expert": true, "enable_balance_scheduling": true,"enable_mlapo":true}' \
     ```
 
 === "A3 series"
@@ -812,7 +810,6 @@ Parameter descriptions:
         export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
         export TASK_QUEUE_ENABLE=1
 
-        export VLLM_ASCEND_ENABLE_MLAPO=1
         export DYNAMIC_EPLB="true"
 
         export ASCEND_RT_VISIBLE_DEVICES=$1
@@ -853,7 +850,7 @@ Parameter descriptions:
                         "ascend_local_comm_res_path": "/etc/hixlep"
                 }
             }' \
-        --additional-config '{"enable_cpu_binding":"True","multistream_overlap_shared_expert":false,"enable_shared_expert_dp":true, "eplb_config":{"dynamic_eplb": true, "expert_heat_collection_interval": 50, "algorithm_execution_interval": 5, "eplb_policy_type": 2, "num_redundant_experts":16}}'
+        --additional-config '{"enable_cpu_binding":"True","multistream_overlap_shared_expert":false,"enable_shared_expert_dp":true,"enable_mlapo":true,"eplb_config":{"dynamic_eplb": true, "expert_heat_collection_interval": 50, "algorithm_execution_interval": 5, "eplb_policy_type": 2, "num_redundant_experts":16}}'
         ```
 
     === "Decode Node"
@@ -877,7 +874,6 @@ Parameter descriptions:
         export OMP_NUM_THREADS=10
         export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
         export TASK_QUEUE_ENABLE=1
-        export VLLM_ASCEND_ENABLE_MLAPO=1
         export ASCEND_RT_VISIBLE_DEVICES=$1
 
         vllm serve /weight/dsk_v3.1-T-w8a8c8_attn-0506-full/ \
@@ -918,13 +914,13 @@ Parameter descriptions:
                         "ascend_local_comm_res_path": "/etc/hixlep"
                 }
             }' \
-        --additional_config '{"enable_cpu_binding": "True", "recompute_scheduler_enable": true, "multistream_overlap_shared_expert": true, "finegrained_tp_config": {"lmhead_tensor_parallel_size":8}}'
+        --additional_config '{"enable_cpu_binding": "True", "recompute_scheduler_enable": true, "multistream_overlap_shared_expert": true, "enable_mlapo": true, "finegrained_tp_config": {"lmhead_tensor_parallel_size":8}}'
         ```
 
     Key Parameter Descriptions:
 
         - `VLLM_ASCEND_ENABLE_FLASHCOMM1=1`: enables the communication optimization function on the prefill nodes.
-        - `VLLM_ASCEND_ENABLE_MLAPO=1`: enables the fusion operator, which can significantly improve performance but consumes more NPU memory. In the Prefill-Decode (PD) separation scenario, enable MLAPO only on decode nodes.
+        - `additional_config.enable_mlapo=true`: enables the fusion operator, which can significantly improve performance but consumes more NPU memory. In the Prefill-Decode (PD) separation scenario, enable MLAPO only on decode nodes.
         - `recompute_scheduler_enable: true`: enables the recomputation scheduler. When the Key-Value Cache (KV Cache) of the decode node is insufficient, requests will be sent to the prefill node to recompute the KV Cache. In the PD separation scenario, it is recommended to enable this configuration on both prefill and decode nodes simultaneously.
         - `multistream_overlap_shared_expert: true`: When the Tensor Parallelism (TP) size is 1 or `enable_shared_expert_dp: true`, an additional stream is enabled to overlap the computation process of shared experts for improved efficiency.
         - `lmhead_tensor_parallel_size: 16`: When the Tensor Parallelism (TP) size of the decode node is 1, this parameter allows the TP size of the LMHead embedding layer to be greater than 1, which is used to reduce the computational load of each card on the LMHead embedding layer.

docs/source/tutorials/models/DeepSeek-V3.2.md

@@ -131,11 +131,10 @@ export OMP_PROC_BIND=false
 export OMP_NUM_THREADS=10
 export VLLM_USE_V1=1
 export HCCL_BUFFSIZE=200
-export VLLM_ASCEND_ENABLE_MLAPO=1
 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
 export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
 
-vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
+vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 --additional-config '{"enable_mlapo":true}' \
 --host 0.0.0.0 \
 --port 8000 \
 --data-parallel-size 2 \
@@ -184,11 +183,10 @@ Run the following scripts on two nodes respectively.
     export OMP_NUM_THREADS=10
     export VLLM_USE_V1=1
     export HCCL_BUFFSIZE=200
-    export VLLM_ASCEND_ENABLE_MLAPO=1
     export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
     export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
 
-    vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
+    vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 --additional-config '{"enable_mlapo":true}' \
     --host 0.0.0.0 \
     --port 8077 \
     --data-parallel-size 2 \
@@ -231,11 +229,10 @@ Run the following scripts on two nodes respectively.
     export OMP_NUM_THREADS=10
     export VLLM_USE_V1=1
     export HCCL_BUFFSIZE=200
-    export VLLM_ASCEND_ENABLE_MLAPO=1
     export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
     export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
 
-    vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
+    vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 --additional-config '{"enable_mlapo":true}' \
     --host 0.0.0.0 \
     --port 8077 \
     --headless \
@@ -282,14 +279,13 @@ Run the following scripts on two nodes respectively.
     export OMP_NUM_THREADS=100
     export VLLM_USE_V1=1
     export HCCL_BUFFSIZE=200
-    export VLLM_ASCEND_ENABLE_MLAPO=1
     export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
     export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
     export HCCL_CONNECT_TIMEOUT=120
     export HCCL_INTRA_PCIE_ENABLE=1
     export HCCL_INTRA_ROCE_ENABLE=0
 
-    vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
+    vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 --additional-config '{"enable_mlapo":true}' \
     --host 0.0.0.0 \
     --port 8077 \
     --data-parallel-size 2 \
@@ -333,14 +329,13 @@ Run the following scripts on two nodes respectively.
     export OMP_NUM_THREADS=100
     export VLLM_USE_V1=1
     export HCCL_BUFFSIZE=200
-    export VLLM_ASCEND_ENABLE_MLAPO=1
     export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
     export VLLM_ASCEND_ENABLE_FLASHCOMM1=1
     export HCCL_CONNECT_TIMEOUT=120
     export HCCL_INTRA_PCIE_ENABLE=1
     export HCCL_INTRA_ROCE_ENABLE=0
 
-    vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 \
+    vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V3.2-W8A8 --additional-config '{"enable_mlapo":true}' \
     --host 0.0.0.0 \
     --port 8077 \
     --headless \

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

@@ -1110,7 +1110,6 @@ export HCCL_SOCKET_IFNAME=$nic_name
 export OMP_PROC_BIND=false
 export OMP_NUM_THREADS=10
 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
-export VLLM_ASCEND_ENABLE_MLAPO=1
 export HCCL_BUFFSIZE=2560
 export TASK_QUEUE_ENABLE=1
 export HCCL_OP_EXPANSION_MODE="AIV"
@@ -1182,7 +1181,7 @@ vllm serve /root/.cache/modelscope/hub/models/vllm-ascend/GLM-5.2-w4a8c8 \
     '{
         "cudagraph_mode": "FULL_DECODE_ONLY",
     }' \
-    --additional-config '{"multistream_overlap_shared_expert": true, "enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true, "recompute_scheduler_enable": true}' \
+    --additional-config '{"multistream_overlap_shared_expert": true, "enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true, "recompute_scheduler_enable": true, "enable_mlapo": true}' \
     --speculative-config '{"num_speculative_tokens": 3, "method":"deepseek_mtp", "enforce_eager":true}'
 ```
 
@@ -1249,7 +1248,7 @@ This 8-node A2 layout splits the global P/D topology across 8 Atlas 800 A2 nodes
 
 **Decode node-specific configurations (A2):**
 
-- `VLLM_ASCEND_ENABLE_MLAPO=1`: MLAPO fusion on decode nodes for memory-bandwidth-bound token generation.
+- `additional_config.enable_mlapo=true`: MLAPO fusion on decode nodes for memory-bandwidth-bound token generation.
 - `--max-num-batched-tokens 256`: Small batch token limit on decode nodes.
 
 **Multi-connector KV transfer configuration (`--kv-transfer-config`):**
@@ -1674,7 +1673,7 @@ The tables below provide recommended parameter configurations for different depl
 |`enable_sparse_sfa_c8` / `enable_sparse_li_c8`|`false` / `true`|`true` / `true` (PD)|`false` / `true`|`enable_sparse_sfa_c8` is an experimental SFA optimization for the C8 quantized model; do not combine it with DCP in v0.23.0 because of known issues. `enable_sparse_li_c8` is independent and remains recommended.|
 |`enable_balance_scheduling`|Enable (single-node)|Enable|Disable in PD mode|Improves output throughput and reduces TPOT in the v1 scheduler. TTFT may degrade in some scenarios; not recommended when Prefill-Decode is separated.|
 |`VLLM_ASCEND_ENABLE_FLASHCOMM1`|1|1 (prefill nodes)|1 (prefill nodes)|Communication optimization. Conflicts with `layer_sharding` containing `o_proj`.|
-|`VLLM_ASCEND_ENABLE_MLAPO`|—|1 (A2 P/D nodes)|—|Fusion operator that significantly improves performance but consumes more NPU memory. On A3 used on decode nodes only.|
+|`additional_config.enable_mlapo`|—|1 (A2 P/D nodes)|—|Fusion operator that significantly improves performance but consumes more NPU memory. On A3 used on decode nodes only.|
 |`cudagraph_mode`|FULL_DECODE_ONLY|FULL_DECODE_ONLY (decode)|FULL_DECODE_ONLY (decode)|Graph capture for the decode phase only. Prefill nodes in PD mode use `--enforce-eager` instead.|
 
 ### 9.2 Tuning Guidelines