Revert "[Feature][MRV2] expand default MRv2 architecture whitelist and add dspark"

tests/e2e/nightly/multi_node/external_dp/config/QWEN3_235B_PD.yaml

@@ -68,7 +68,8 @@ env_decode: &env_decode
   DISABLE_L2_CACHE: "1"
   ASCEND_RT_VISIBLE_DEVICES: "${VISIBLE_DEVICES}"
   VLLM_TORCH_PROFILER_WITH_STACK: 0
-  VLLM_USE_V2_MODEL_RUNNER: "1"
+  EXPERT_MAP_RECORD: "true"
+  DYNAMIC_EPLB: "true"
   HCCL_BUFFSIZE: "1024"
 
 templates:
@@ -147,12 +148,9 @@ templates:
       - --async-scheduling
       - --compilation-config
       - '{"cudagraph_mode": "FULL_DECODE_ONLY","cudagraph_capture_sizes":[6,12,18]}'
-      - --enable-eplb
-      - --eplb-config
-      - '{"window_size":4000,"step_interval":200,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
       - --additional-config
-      # Model Runner V2 keeps only the Ascend load-collection phase here.
-      - '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"eplb_config":{"load_collection_phase":"all"},"enable_mlapo":true}'
+      # #5533 renamed and nested the legacy EPLB keys below.
+      - '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"eplb_config":{"dynamic_eplb":true,"expert_heat_collection_interval":4000,"algorithm_execution_interval":200,"num_redundant_experts":0},"enable_mlapo":true}'
       - --kv-transfer-config
       - '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "1", "kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 1, "tp_size": 16}, "decode": {"dp_size": 32, "tp_size": 1}}}'
 
@@ -192,12 +190,9 @@ templates:
       - --async-scheduling
       - --compilation-config
       - '{"cudagraph_mode": "FULL_DECODE_ONLY","cudagraph_capture_sizes":[6,12,18]}'
-      - --enable-eplb
-      - --eplb-config
-      - '{"window_size":4000,"step_interval":200,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
       - --additional-config
-      # Model Runner V2 keeps only the Ascend load-collection phase here.
-      - '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"eplb_config":{"load_collection_phase":"all"},"enable_mlapo":true}'
+      # #5533 renamed and nested the legacy EPLB keys below.
+      - '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"eplb_config":{"dynamic_eplb":true,"expert_heat_collection_interval":4000,"algorithm_execution_interval":200,"num_redundant_experts":0},"enable_mlapo":true}'
       - --kv-transfer-config
       - '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "1", "kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 1, "tp_size": 16}, "decode": {"dp_size": 32, "tp_size": 1}}}'
 

tests/e2e/nightly/multi_node/external_dp/config/QWEN3_235B_PD_3_5K_1_5k.yaml

@@ -66,7 +66,8 @@ env_prefill: &env_prefill
 env_decode: &env_decode
   <<: *env_common
   VLLM_TORCH_PROFILER_WITH_STACK: 0
-  VLLM_USE_V2_MODEL_RUNNER: "1"
+  EXPERT_MAP_RECORD: "true"
+  DYNAMIC_EPLB: "true"
   HCCL_BUFFSIZE: "1024"
 
 templates:
@@ -146,12 +147,9 @@ templates:
       - --no-enable-prefix-caching
       - --compilation-config
       - '{"cudagraph_mode": "FULL_DECODE_ONLY","cudagraph_capture_sizes":[5,10,15]}'
-      - --enable-eplb
-      - --eplb-config
-      - '{"window_size":4000,"step_interval":200,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
       - --additional-config
-      # Model Runner V2 keeps only the Ascend load-collection phase here.
-      - '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"eplb_config":{"load_collection_phase":"all"},"enable_mlapo":true}'
+      # #5533 renamed and nested the legacy EPLB keys below.
+      - '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"eplb_config":{"dynamic_eplb":true,"expert_heat_collection_interval":4000,"algorithm_execution_interval":200,"num_redundant_experts":0},"enable_mlapo":true}'
       - --kv-transfer-config
       - '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
 
@@ -190,12 +188,9 @@ templates:
       - --no-enable-prefix-caching
       - --compilation-config
       - '{"cudagraph_mode": "FULL_DECODE_ONLY","cudagraph_capture_sizes":[5,10,15]}'
-      - --enable-eplb
-      - --eplb-config
-      - '{"window_size":4000,"step_interval":200,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
       - --additional-config
-      # Model Runner V2 keeps only the Ascend load-collection phase here.
-      - '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"eplb_config":{"load_collection_phase":"all"},"enable_mlapo":true}'
+      # #5533 renamed and nested the legacy EPLB keys below.
+      - '{"finegrained_tp_config": {"lmhead_tensor_parallel_size":8},"eplb_config":{"dynamic_eplb":true,"expert_heat_collection_interval":4000,"algorithm_execution_interval":200,"num_redundant_experts":0},"enable_mlapo":true}'
       - --kv-transfer-config
       - '{"kv_connector": "MooncakeConnectorV1", "kv_role": "kv_consumer", "kv_port": "30200", "engine_id": "2", "kv_connector_extra_config": {"use_ascend_direct": true,"prefill": {"dp_size": 2, "tp_size": 8}, "decode": {"dp_size": 32, "tp_size": 1}}}'
 

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

@@ -12,7 +12,7 @@ env_common: &env_common
   HCCL_DETERMINISTIC: True
   TASK_QUEUE_ENABLE: 1
   HCCL_OP_RETRY_ENABLE: "L0:0, L1:0, L2:0"
-  VLLM_USE_V2_MODEL_RUNNER: "1"
+  DYNAMIC_EPLB: true
   VLLM_ENGINE_READY_TIMEOUT_S: 3000
 
 disaggregated_prefill:
@@ -56,10 +56,8 @@ deployment:
                     }
               }
           }'
-          --enable-eplb
-          --eplb-config '{"window_size":2048,"step_interval":200,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
           --additional-config
-          '{"weight_nz_mode":1,"eplb_config":{"load_collection_phase":"all"}}'
+          '{"weight_nz_mode":1,"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":2048,"algorithm_execution_interval":200}}'
 
   -
     envs:
@@ -96,13 +94,12 @@ deployment:
                     }
               }
           }'
-          --enable-eplb
-          --eplb-config '{"window_size":2048,"step_interval":200,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
           --additional-config
-          '{"weight_nz_mode":1,"eplb_config":{"load_collection_phase":"all"}}'
+          '{"weight_nz_mode":1,"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":2048,"algorithm_execution_interval":200}}'
   -
     envs:
       <<: *env_common
+    # #5533 requires the EPLB keys in the command below to be nested.
     server_cmd: >
       vllm serve vllm-ascend/DeepSeek-R1-0528-W8A8
         --host 0.0.0.0
@@ -136,10 +133,8 @@ deployment:
                   }
             }
         }'
-        --enable-eplb
-        --eplb-config '{"window_size":2048,"step_interval":200,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
         --additional-config
-        '{"multistream_overlap_shared_expert":true,"eplb_config":{"load_collection_phase":"all"}}'
+        '{"multistream_overlap_shared_expert":true,"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":2048,"algorithm_execution_interval":200}}'
   -
     envs:
       <<: *env_common
@@ -176,10 +171,8 @@ deployment:
                   }
             }
         }'
-        --enable-eplb
-        --eplb-config '{"window_size":2048,"step_interval":200,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
         --additional-config
-        '{"multistream_overlap_shared_expert":true,"eplb_config":{"load_collection_phase":"all"}}'
+        '{"multistream_overlap_shared_expert":true,"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":2048,"algorithm_execution_interval":200}}'
 benchmarks:
   perf:
     case_type: performance

tests/e2e/nightly/multi_node/internal_dp/config/GLM5_1-W8A8-A3-dual-nodes.yaml

@@ -79,17 +79,7 @@ benchmarks:
     batch_size: 32
     baseline: 90
     threshold: 10
-  perf_warmup:
-    case_type: performance
-    dataset_path: vllm-ascend/GSM8K_prefix90_in131072_bs100_glm
-    request_conf: vllm_api_stream_chat
-    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
-    num_prompts: 8
-    max_out_len: 1024
-    batch_size: 2
-    request_rate: 0
-    baseline: 0
-    threshold: 0.97
+
   perf:
     case_type: performance
     dataset_path: vllm-ascend/GSM8K_prefix90_in131072_bs100_glm

tests/e2e/nightly/multi_node/internal_dp/config/GLM5_2-W8A8-A3-dual-nodes.yaml

@@ -73,17 +73,7 @@ benchmarks:
     batch_size: 32
     baseline: 90
     threshold: 10
-  perf_warmup:
-    case_type: performance
-    dataset_path: vllm-ascend/GSM8K_prefix90_in131072_bs100_glm
-    request_conf: vllm_api_stream_chat
-    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
-    num_prompts: 8
-    max_out_len: 1024
-    batch_size: 2
-    request_rate: 0
-    baseline: 0
-    threshold: 0.97
+
   perf:
     case_type: performance
     dataset_path: vllm-ascend/GSM8K_prefix90_in131072_bs100_glm

tests/e2e/nightly/multi_node/internal_dp/config/Qwen3-235B-W8A8-EPLB-mxfp0.yaml

@@ -11,7 +11,7 @@ env_common: &env_common
   OMP_NUM_THREADS: 1
   HCCL_BUFFSIZE: 1024
   SERVER_PORT: 8080
-  VLLM_USE_V2_MODEL_RUNNER: "1"
+  DYNAMIC_EPLB: true
 disaggregated_prefill:
   enabled: true
   prefiller_host_index: [0]
@@ -52,10 +52,8 @@ deployment:
                   }
             }
         }'
-        --enable-eplb
-        --eplb-config '{"window_size":50,"step_interval":5,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
         --additional-config
-        '{"eplb_config":{"load_collection_phase":"all"}}'
+        '{"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":50,"algorithm_execution_interval":5}}'
 
   -
     envs:
@@ -91,8 +89,6 @@ deployment:
                   }
             }
         }'
-        --enable-eplb
-        --eplb-config '{"window_size":600,"step_interval":50,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
         --additional-config
-        '{"eplb_config":{"load_collection_phase":"all"}}'
+        '{"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":600,"algorithm_execution_interval":50}}'
 benchmarks:

tests/e2e/nightly/multi_node/internal_dp/config/Qwen3-235B-W8A8-EPLB.yaml

@@ -10,7 +10,7 @@ env_common: &env_common
   OMP_NUM_THREADS: 1
   HCCL_BUFFSIZE: 1024
   SERVER_PORT: 8080
-  VLLM_USE_V2_MODEL_RUNNER: "1"
+  DYNAMIC_EPLB: true
 disaggregated_prefill:
   enabled: true
   prefiller_host_index: [0]
@@ -51,10 +51,8 @@ deployment:
                   }
             }
         }'
-        --enable-eplb
-        --eplb-config '{"window_size":50,"step_interval":5,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
         --additional-config
-        '{"eplb_config":{"load_collection_phase":"all"}}'
+        '{"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":50,"algorithm_execution_interval":5}}'
 
   -
     envs:
@@ -90,8 +88,6 @@ deployment:
                   }
             }
         }'
-        --enable-eplb
-        --eplb-config '{"window_size":600,"step_interval":50,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
         --additional-config
-        '{"eplb_config":{"load_collection_phase":"all"}}'
+        '{"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":600,"algorithm_execution_interval":50}}'
 benchmarks:

tests/e2e/nightly/single_node/models/configs/DeepSeek-V4-Flash-W8A8-A3.yaml

@@ -63,7 +63,6 @@ test_cases:
         baseline: 86.36
         threshold: 5
         thinking: true
-        reasoning_effort: low
       perf:
         case_type: performance
         dataset_path: vllm-ascend/GSM8K_prefix0_in32768_bs1000_deepseek

tests/e2e/nightly/single_node/models/configs/GLM5_1_W4A4_A5.yaml

@@ -18,7 +18,7 @@ _envs: &envs
 #  HCCL_INTRA_PCIE_ENABLE: "1"
   HCCL_INTRA_ROCE_ENABLE: "0"
   ASCEND_LAUNCH_BLOCKING: "0"
-  VLLM_USE_V2_MODEL_RUNNER: "1"
+  DYNAMIC_EPLB: "true"
 
 _server_cmd: &server_cmd
   - "--port"
@@ -32,9 +32,6 @@ _server_cmd: &server_cmd
   - "--seed"
   - "1024"
   - "--enable-expert-parallel"
-  - "--enable-eplb"
-  - "--eplb-config"
-  - '{"window_size":50,"step_interval":5,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
   - "--async-scheduling"
   - "--max-num-seqs"
   - "128"
@@ -60,7 +57,7 @@ _server_cmd: &server_cmd
   - "--speculative-config"
   - '{"num_speculative_tokens":3,"method":"deepseek_mtp","enforce_eager":true}'
   - "--additional-config"
-  - '{"enable_cpu_binding":"True","multistream_overlap_shared_expert":"True","enable_sparse_sfa_c8":true,"enable_sparse_li_c8":true,"enable_dsa_cp":true,"eplb_config":{"load_collection_phase":"all"}}'
+  - '{"enable_cpu_binding":"True","multistream_overlap_shared_expert":"True","enable_sparse_sfa_c8":true,"enable_sparse_li_c8":true,"enable_dsa_cp":true,"eplb_config":{"dynamic_eplb":true,"expert_heat_collection_interval":50,"algorithm_execution_interval":5,"eplb_policy_type":2,"num_redundant_experts":0}}'
 
 # ==========================================
 # ACTUAL TEST CASES

tests/e2e/nightly/single_node/models/configs/Kimi-K2.6-w4a8-A3.yaml

@@ -12,7 +12,7 @@ test_cases:
       TASK_QUEUE_ENABLE: "1"
       HCCL_BUFFSIZE: "800"
       VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
-      VLLM_USE_V2_MODEL_RUNNER: "1"
+      DYNAMIC_EPLB: "true"
       SERVER_PORT: "DEFAULT_PORT"
     server_cmd:
       - "--quantization"
@@ -30,9 +30,6 @@ test_cases:
       - "2"
       - "--no-enable-prefix-caching"
       - "--enable-expert-parallel"
-      - "--enable-eplb"
-      - "--eplb-config"
-      - '{"window_size":600,"step_interval":50,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
       - "--max-num-seqs"
       - "24"
       - "--max-model-len"
@@ -47,7 +44,7 @@ test_cases:
       - "--compilation-config"
       - '{"cudagraph_mode":"FULL_DECODE_ONLY"}'
       - "--additional-config"
-      - '{"eplb_config":{"load_collection_phase":"all"},"ascend_compilation_config":{"enable_static_kernel":false},"enable_fused_mc2":1,"scheduler_config":{"enable_balance_scheduling":true},"enable_mlapo":true}'
+      - '{"eplb_config":{"dynamic_eplb":true},"ascend_compilation_config":{"enable_static_kernel":false},"enable_fused_mc2":1,"scheduler_config":{"enable_balance_scheduling":true},"enable_mlapo":true}'
       - "--profiler-config"
       - '{"profiler": "torch", "torch_profiler_dir": "./vllm_profile", "torch_profiler_with_stack": true}'
       - "--mm-processor-cache-gb"

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

@@ -72,13 +72,11 @@ test_cases:
     model: "vllm-ascend/Qwen3-235B-A22B-W8A8"
     envs:
       <<: *envs
+      DYNAMIC_EPLB: "true"
     server_cmd: *server_cmd
     server_cmd_extra:
-      - "--enable-eplb"
-      - "--eplb-config"
-      - '{"window_size": 600, "step_interval": 50, "num_redundant_experts": 16, "use_async": true, "log_balancedness": true, "log_balancedness_interval": 1}'
       - "--additional-config"
-      - '{"eplb_config": {"load_collection_phase": "all"}}'
+      - '{"eplb_config": {"dynamic_eplb": "true", "expert_heat_collection_interval": 600, "algorithm_execution_interval": 50, "num_redundant_experts": 16, "eplb_policy_type": 2}}'
       - "--compilation-config"
       - '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
     benchmarks:

tests/e2e/nightly/single_node/models/configs/Qwen3-235B-A22B-W8A8_piecewise_fullgraph_A3.yaml

@@ -94,14 +94,9 @@ test_cases:
     model: "Eco-Tech/Qwen3-235B-A22B-w8a8-QuaRot"
     envs:
       <<: *envs
-      VLLM_USE_V2_MODEL_RUNNER: "1"
+      DYNAMIC_EPLB: "true"
     server_cmd: *server_cmd
     server_cmd_extra:
-      - "--enable-eplb"
-      - "--eplb-config"
-      - '{"window_size":600,"step_interval":50,"num_redundant_experts":0,"use_async":true,"log_balancedness":true,"log_balancedness_interval":1}'
-      - "--additional-config"
-      - '{"eplb_config":{"load_collection_phase":"all"}}'
       - "--compilation-config"
       - '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
     benchmarks:

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

@@ -94,10 +94,6 @@ def server(
         **os.environ,
         "VLLM_SERVER_DEV_MODE": "1",
         "HF_HUB_OFFLINE": "1",
-        # Qwen3ForCausalLM now defaults to MRv2. Sleep/wake generate is still
-        # V1-only; pin the RLHF server to V1 until the MRv2 allocator path is
-        # ready. Keep this on the subprocess env only.
-        "VLLM_USE_V2_MODEL_RUNNER": "0",
     }
     base = _DUMMY_ARGS if dummy_weights else _BASE_ARGS
     cmd = [

tests/e2e/pull_request/one_card/spec_decode/utils.py

@@ -42,12 +42,10 @@
     "dspark": [1.0, 0.8, 0.6, 0.6, 0.6, 0.6, 0.6],
     "dspark_dynamic": [0.68, 0.54, 0.46, 0.38, 0.22, 0.11, 0.11],
     # DSpark with draft_window_size=512 on a prompt longer than the window
-    # (test_dspark_kv_sliding_window in test_dspark.py). MRv2 blacklists
-    # DSpark KV sliding window, so this case stays on the V1 runner. Measured
-    # on the deepseek-ai/dspark_qwen3_8b_block7 single-block draft in CI;
-    # positions 3-6 exceed the no-window "dspark" baseline because the
-    # single-block drafter relies on the most recent context, which the
-    # window keeps.
+    # (test_dspark_kv_sliding_window in test_dspark.py). Measured on the
+    # deepseek-ai/dspark_qwen3_8b_block7 single-block draft in CI; positions
+    # 3-6 exceed the no-window "dspark" baseline because the single-block
+    # drafter relies on the most recent context, which the window keeps.
     "dspark_sliding_window": [0.88, 0.88, 0.88, 0.88, 0.75, 0.75, 0.62],
 }
 

tests/e2e/pull_request/one_card/test_attention_v1_precision.py

@@ -21,13 +21,7 @@
 
 
 @pytest.fixture(autouse=True)
-def default_vllm_config(monkeypatch):
-    # Qwen3 now defaults to the V2 model runner, whose forward-context hook
-    # (NPUPlatform.set_additional_forward_context) queries TP/DP groups. This
-    # suite drives attention backends in a single process with no engine, so
-    # mock the group accessors, mirroring patch_distributed_groups above.
-    monkeypatch.setattr("vllm.distributed.get_tensor_model_parallel_world_size", lambda: 1)
-    monkeypatch.setattr("vllm.distributed.get_dp_group", lambda: MagicMock(world_size=1))
+def default_vllm_config():
     mock_config = MagicMock()
     mock_config.compilation_config = MagicMock()
     mock_config.compilation_config.custom_ops = ["all"]