Default enable MLAPO (#5952)

### What this PR does / why we need it?
1) Default enable MLAPO for deepseek MLA Attention W8A8 models on PD
disagregation D Instance, for example: DeepSeekV3-W8A8,
DeepSeek-R1-W8A8.
2) Default enable MLAPO for DeepSeek SFA Attention W8A8 models,
currently is DeepSeek-V3.2-W8A8.

### Does this PR introduce _any_ user-facing change?
Don't need use manully to VLLM_ASCEND_ENABLE_MLAPO=1, to enable MLAPO
feature for deepseek w8a8 model

The effect of enabling MLAPO SFA model deployed on a single A3 Node:
Test
with:tests/e2e/nightly/single_node/models/test_deepseek_v3_2_exp_w8a8.py
dataset: gsm8k-lite，without set MTP, FULL GRAPH, has 19% promote：
未默认开启 MLAPO 时：
├─────────────────────────┤
│                TTFT                      │ 14055.8836 ms   │
├─────────────────────────┤
│                ITL                         │ 66.8171 ms.          │
├─────────────────────────┤
│ Output Token Throughput  │ 104.9105 token/s │
├─────────────────────────┤
默认开启 MLAPO 时：
├─────────────────────────┤
│                TTFT                      │ 3753.1547 ms   │
├─────────────────────────┤
│                ITL.                        │ 61.4236  ms.       │
├─────────────────────────┤
│ Output Token Throughput  │ 125.2075 token/s│
├─────────────────────────┤

- vLLM version: v0.13.0
- vLLM main:
2c24bc6996

---------

Signed-off-by: leo-pony <nengjunma@outlook.com>

docs/source/tutorials/DeepSeek-R1.md

@@ -94,7 +94,6 @@ 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_MLAPO=1
 export VLLM_ASCEND_BALANCE_SCHEDULING=1
 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
 export VLLM_USE_MODELSCOPE=True
@@ -121,7 +120,6 @@ vllm serve vllm-ascend/DeepSeek-R1-W8A8 \
 **Notice:**
 The parameters are explained as follows:
 
-- Setting the environment variable `VLLM_ASCEND_ENABLE_MLAPO=1` enables a fusion operator that can significantly improve performance, though it requires more NPU memory. It is therefore recommended to enable this option when sufficient NPU memory is available.
 - Setting the environment variable `VLLM_ASCEND_BALANCE_SCHEDULING=1` enables balance scheduling. This may help increase output throughput and reduce TPOT in v1 scheduler. However, TTFT may degrade in some scenarios. Furthermore, enabling this feature is not recommended in scenarios where PD is separated.
 - For single-node deployment, we recommend using `dp4tp4` instead of `dp2tp8`.
 - `--max-model-len` specifies the maximum context length - that is, the sum of input and output tokens for a single request. For performance testing with an input length of 3.5K and output length of 1.5K, a value of `16384` is sufficient, however, for precision testing, please set it at least `35000`.
@@ -151,7 +149,6 @@ export OMP_PROC_BIND=false
 export OMP_NUM_THREADS=1
 export HCCL_BUFFSIZE=200
 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
-export VLLM_ASCEND_ENABLE_MLAPO=1
 export VLLM_ASCEND_BALANCE_SCHEDULING=1
 export HCCL_INTRA_PCIE_ENABLE=1
 export HCCL_INTRA_ROCE_ENABLE=0
@@ -198,7 +195,6 @@ export OMP_PROC_BIND=false
 export OMP_NUM_THREADS=1
 export HCCL_BUFFSIZE=200
 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
-export VLLM_ASCEND_ENABLE_MLAPO=1
 export VLLM_ASCEND_BALANCE_SCHEDULING=1
 export HCCL_INTRA_PCIE_ENABLE=1
 export HCCL_INTRA_ROCE_ENABLE=0

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

@@ -106,7 +106,6 @@ 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_MLAPO=1
 export VLLM_ASCEND_BALANCE_SCHEDULING=1
 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
 
@@ -133,7 +132,6 @@ vllm serve /weights/DeepSeek-V3.1-w8a8-mtp-QuaRot \
 **Notice:**
 The parameters are explained as follows:
 
-- Setting the environment variable `VLLM_ASCEND_ENABLE_MLAPO=1` enables a fusion operator that can significantly improve performance, though it requires more NPU memory. It is therefore recommended to enable this option when sufficient NPU memory is available.
 - Setting the environment variable `VLLM_ASCEND_BALANCE_SCHEDULING=1` enables balance scheduling. This may help increase output throughput and reduce TPOT in v1 scheduler. However, TTFT may degrade in some scenarios. Furthermore, enabling this feature is not recommended in scenarios where PD is separated.
 - For single-node deployment, we recommend using `dp4tp4` instead of `dp2tp8`.
 - `--max-model-len` specifies the maximum context length - that is, the sum of input and output tokens for a single request. For performance testing with an input length of 3.5K and output length of 1.5K, a value of `16384` is sufficient, however, for precision testing, please set it at least `35000`.
@@ -171,7 +169,6 @@ export OMP_PROC_BIND=false
 export OMP_NUM_THREADS=1
 export HCCL_BUFFSIZE=200
 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
-export VLLM_ASCEND_ENABLE_MLAPO=1
 export VLLM_ASCEND_BALANCE_SCHEDULING=1
 export HCCL_INTRA_PCIE_ENABLE=1
 export HCCL_INTRA_ROCE_ENABLE=0
@@ -224,7 +221,6 @@ export OMP_PROC_BIND=false
 export OMP_NUM_THREADS=1
 export HCCL_BUFFSIZE=200
 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
-export VLLM_ASCEND_ENABLE_MLAPO=1
 export VLLM_ASCEND_BALANCE_SCHEDULING=1
 export HCCL_INTRA_PCIE_ENABLE=1
 export HCCL_INTRA_ROCE_ENABLE=0
@@ -449,7 +445,6 @@ export HCCL_CONNECT_TIMEOUT=120
 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=1100
 export TASK_QUEUE_ENABLE=1
 export HCCL_OP_EXPANSION_MODE="AIV"
@@ -526,7 +521,6 @@ export HCCL_CONNECT_TIMEOUT=120
 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=1100
 export TASK_QUEUE_ENABLE=1
 export HCCL_OP_EXPANSION_MODE="AIV"

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

@@ -441,8 +441,6 @@ Before you start, please
 
         export ASCEND_RT_VISIBLE_DEVICES=$1
 
-        export VLLM_ASCEND_ENABLE_MLAPO=1
-
 
         vllm serve /root/.cache/Eco-Tech/DeepSeek-V3.2-w8a8-mtp-QuaRot \
             --host 0.0.0.0 \
@@ -520,8 +518,6 @@ Before you start, please
 
         export ASCEND_RT_VISIBLE_DEVICES=$1
 
-        export VLLM_ASCEND_ENABLE_MLAPO=1
-
 
         vllm serve /root/.cache/Eco-Tech/DeepSeek-V3.2-w8a8-mtp-QuaRot \
             --host 0.0.0.0 \

docs/source/tutorials/long_sequence_context_parallel_multi_node.md

@@ -224,7 +224,6 @@ export OMP_NUM_THREADS=1
 export HCCL_OP_EXPANSION_MODE="AIV"
 export VLLM_USE_V1=1
 export TASK_QUEUE_ENABLE=1
-export VLLM_ASCEND_ENABLE_MLAPO="1"
 export VLLM_ASCEND_ENABLE_CONTEXT_PARALLEL=1
 
 vllm serve /path_to_weight/DeepSeek-V3.1_w8a8mix_mtp \

docs/source/tutorials/pd_disaggregation_mooncake_multi_node.md

@@ -370,7 +370,6 @@ export HCCL_SOCKET_IFNAME=$nic_name
 export OMP_PROC_BIND=false
 export OMP_NUM_THREADS=10
 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
-export VLLM_ASCEND_ENABLE_MLAPO=1
 export HCCL_BUFFSIZE=600
 export TASK_QUEUE_ENABLE=1
 export HCCL_OP_EXPANSION_MODE="AIV"
@@ -429,7 +428,6 @@ export HCCL_SOCKET_IFNAME=$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=600
 export TASK_QUEUE_ENABLE=1
 export HCCL_OP_EXPANSION_MODE="AIV"
@@ -615,7 +613,6 @@ export HCCL_SOCKET_IFNAME=$nic_name
 export OMP_PROC_BIND=false
 export OMP_NUM_THREADS=10
 export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
-export VLLM_ASCEND_ENABLE_MLAPO=1
 export HCCL_BUFFSIZE=600
 export TASK_QUEUE_ENABLE=1
 export HCCL_OP_EXPANSION_MODE="AIV"
@@ -674,7 +671,6 @@ export HCCL_SOCKET_IFNAME=$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=600
 export TASK_QUEUE_ENABLE=1
 export HCCL_OP_EXPANSION_MODE="AIV"

tests/e2e/nightly/multi_node/config/DeepSeek-R1-W8A8-A2.yaml

@@ -3,7 +3,6 @@ model: "vllm-ascend/DeepSeek-R1-0528-W8A8"
 num_nodes: 2
 npu_per_node: 8
 env_common:
-  VLLM_ASCEND_ENABLE_MLAPO: 1
   VLLM_ASCEND_BALANCE_SCHEDULING: 1
   HCCL_INTRA_PCIE_ENABLE: 1
   HCCL_INTRA_ROCE_ENABLE: 0

tests/e2e/nightly/multi_node/config/DeepSeek-V3_2-W8A8-A3-dual-nodes.yaml

@@ -10,7 +10,6 @@ env_common:
   SERVER_PORT: 8080
   OMP_PROC_BIND: false
   OMP_NUM_THREADS: 1
-  VLLM_ASCEND_ENABLE_MLAPO: 1
   PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
   VLLM_ASCEND_ENABLE_FLASHCOMM1: 0
   ASCEND_A3_EBA_ENABLE: 1

tests/e2e/nightly/single_node/models/test_deepseek_r1_0528_w8a8.py

@@ -31,7 +31,6 @@ MODELS = [
 MODES = [
     "single",
     "aclgraph",
-    "aclgraph_mlapo",
 ]
 
 prompts = [
@@ -88,8 +87,6 @@ async def test_models(model: str, mode: str) -> None:
     ]
     if mode == "single":
         server_args.append("--enforce-eager")
-    if mode == "aclgraph_mlapo":
-        env_dict["VLLM_ASCEND_ENABLE_MLAPO"] = "1"
     server_args.extend(["--additional-config", json.dumps(additional_config)])
     request_keyword_args: dict[str, Any] = {
         **api_keyword_args,

tests/e2e/nightly/single_node/models/test_deepseek_r1_0528_w8a8_eplb.py

@@ -56,7 +56,6 @@ async def test_models(model: str) -> None:
         "OMP_NUM_THREADS": "100",
         "OMP_PROC_BIND": "false",
         "HCCL_BUFFSIZE": "200",
-        "VLLM_ASCEND_ENABLE_MLAPO": "1",
         "VLLM_RPC_TIMEOUT": "3600000",
         "VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS": "3600000",
         "DISABLE_L2_CACHE": "1",

vllm_ascend/attention/mla_v1.py

@@ -22,7 +22,8 @@ from vllm_ascend.attention.context_parallel.common_cp import (
 from vllm_ascend.attention.utils import (
     AscendCommonAttentionMetadata, ascend_chunked_prefill_workspace_size,
     enable_cp, maybe_save_kv_layer_to_connector, split_decodes_and_prefills,
-    trans_rope_weight, transdata, wait_for_kv_layer_from_connector)
+    trans_rope_weight, transdata, wait_for_kv_layer_from_connector,
+    enabling_malpo)
 from vllm_ascend.compilation.acl_graph import (
     get_draft_graph_params, get_graph_params,
     update_draft_graph_params_workspaces, update_graph_params_workspaces)
@@ -741,7 +742,7 @@ class AscendMLAImpl(MLAAttentionImpl):
         self.ring_mla_mask_size = 512
 
         self.speculative_config = self.vllm_config.speculative_config
-        self.enable_mlapo = envs.VLLM_ASCEND_ENABLE_MLAPO
+        self.enable_mlapo = enabling_malpo(self.vllm_config)
 
         self.is_kv_producer = self.vllm_config.kv_transfer_config is not None and self.vllm_config.kv_transfer_config.is_kv_producer
         self.layer_sharding_kwargs = []
@@ -1491,7 +1492,6 @@ class AscendMLAImpl(MLAAttentionImpl):
 
         # MLA Preprocess
         if self.enable_mlapo and \
-            not has_prefill and \
             attn_metadata.num_decode_tokens <= MLAPO_MAX_SUPPORTED_TOKENS:
             hidden_states = torch.ops.vllm.maybe_all_gather_and_maybe_unpad(
                 hidden_states.contiguous(), need_gather_q_kv)

vllm_ascend/attention/sfa_v1.py

@@ -375,6 +375,9 @@ class AscendSFAImpl(MLAAttentionImpl):
         ascend_config = get_ascend_config()
         self.enable_shared_expert_dp = ascend_config.enable_shared_expert_dp
         self.enable_prefetch = ascend_config.weight_prefetch_config.enabled
+
+        # In sfa, prefill and decode have the same calculation formula,
+        # so do not distinguish between prefill and decode here.
         self.enable_mlapo = envs.VLLM_ASCEND_ENABLE_MLAPO
 
         assert self.indexer is not None, "Indexer is required for DSA."

vllm_ascend/attention/utils.py

@@ -9,6 +9,7 @@ from vllm.distributed.kv_transfer import get_kv_transfer_group, has_kv_transfer_
 from vllm.forward_context import ForwardContext, get_forward_context
 from vllm.v1.attention.backends.utils import CommonAttentionMetadata
 
+from vllm_ascend import envs
 from vllm_ascend.utils import AscendDeviceType, get_ascend_config, get_ascend_device_type
 
 
@@ -302,3 +303,8 @@ def transdata(nd_mat, block_size: tuple = (16, 16)):
     )
     nz_mat = torch.reshape(nz_mat, (nz_mat.shape[0], nz_mat.shape[1] * nz_mat.shape[2], nz_mat.shape[3]))
     return nz_mat
+
+
+def enabling_malpo(vllm_config: VllmConfig) -> bool:
+    is_decode_instance = vllm_config.kv_transfer_config is not None and vllm_config.kv_transfer_config.is_kv_consumer
+    return bool(envs.VLLM_ASCEND_ENABLE_MLAPO and is_decode_instance)

vllm_ascend/envs.py

@@ -92,7 +92,11 @@ env_variables: dict[str, Callable[[], Any]] = {
     ),
     # Whether to enable msMonitor tool to monitor the performance of vllm-ascend.
     "MSMONITOR_USE_DAEMON": lambda: bool(int(os.getenv("MSMONITOR_USE_DAEMON", "0"))),
-    "VLLM_ASCEND_ENABLE_MLAPO": lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_MLAPO", "0"))),
+    # Whether to enable MLAPO optimization for DeepSeek W8A8 series models.
+    # This option is enabled by default. MLAPO can improve performance, but
+    # it will consume more NPU memory. If reducing NPU memory usage is a higher priority
+    # for your DeepSeek W8A8 scene, then disable it.
+    "VLLM_ASCEND_ENABLE_MLAPO": lambda: bool(int(os.getenv("VLLM_ASCEND_ENABLE_MLAPO", "1"))),
     # Whether to enable weight cast format to FRACTAL_NZ.
     # 0: close nz;
     # 1: only quant case enable nz;