init v0.23.0

Signed-off-by: Sun Ruoxi <sunruoxi@4paradigm.com>
2026-08-27 15:11:51 +08:00
parent b582a8e7d1
commit 7f8a1b1f7a
2849 changed files with 712887 additions and 22001 deletions
--- a/tests/e2e/weekly/multi_node/internal_dp/config/DeepSeek-V3.yaml
+++ b/tests/e2e/weekly/multi_node/internal_dp/config/DeepSeek-V3.yaml
@@ -0,0 +1,112 @@
+# For disaggregated mode, set is_disaggregated: true, and set the following parameters:
+# Prefiller_index: the hosts index of the node running prefiller
+# Decoder_index: the hosts index of the node running decoder
+# Suppose we have **4 nodes** running a 2P1D setup (2 Prefillers + 1 Decoder):
+#   ┌───────────────┬───────────────┬───────────────┬───────────────┐
+#   │   node0       │   node1       │   node2       │   node3       │
+#   │ Prefiller #1  │ Prefiller #2  │   Decoder     │   Decoder     │
+#   └───────────────┴───────────────┴───────────────┴───────────────┘
+# For the prefiller nodes. the hosts should be node0 and node1
+# For the decoder nodes. we only have 1 decoder node(dp+tp+ep across node2 and node3. Where node3 is running with headless mode)
+# So the prefiller_host_index is [0, 1], and the decoder_host_index is [2]
+test_name: "test DeepSeek-V3 disaggregated_prefill"
+model: "vllm-ascend/DeepSeek-V3-W8A8"
+num_nodes: 2
+npu_per_node: 16
+env_common: &env_common
+  VLLM_USE_MODELSCOPE: true
+  OMP_PROC_BIND: false
+  OMP_NUM_THREADS: 1
+  TASK_QUEUE_ENABLE: 1
+  HCCL_BUFFSIZE: 1024
+  SERVER_PORT: 8080
+  NUMEXPR_MAX_THREADS: 128
+  DISAGGREGATED_PREFILL_PROXY_SCRIPT: "examples/disaggregated_prefill_v1/load_balance_proxy_server_example.py"
+# For None kubernetes deployment, list the IPs of all nodes used in order as follow
+# cluster_hosts: [10.0.0.103, 10.0.0.139]
+disaggregated_prefill:
+  enabled: true
+  prefiller_host_index: [0]
+  decoder_host_index: [1]
+
+deployment:
+  -
+    envs:
+      <<: *env_common
+    server_cmd: >
+        vllm serve "vllm-ascend/DeepSeek-V3-W8A8"
+        --host 0.0.0.0
+        --port $SERVER_PORT
+        --data-parallel-size 2
+        --data-parallel-size-local 2
+        --tensor-parallel-size 8
+        --seed 1024
+        --enforce-eager
+        --enable-expert-parallel
+        --max-num-seqs 16
+        --max-model-len 8192
+        --max-num-batched-tokens 8192
+        --quantization ascend
+        --trust-remote-code
+        --no-enable-prefix-caching
+        --gpu-memory-utilization 0.9
+        --kv-transfer-config
+        '{"kv_connector": "MooncakeConnectorV1",
+        "kv_role": "kv_producer",
+        "kv_port": "30000",
+        "kv_connector_extra_config": {
+                  "prefill": {
+                          "dp_size": 2,
+                          "tp_size": 8
+                  },
+                  "decode": {
+                          "dp_size": 2,
+                          "tp_size": 8
+                  }
+            }
+        }'
+
+  -
+    envs:
+      <<: *env_common
+    server_cmd: >
+        vllm serve "vllm-ascend/DeepSeek-V3-W8A8"
+        --host 0.0.0.0
+        --port $SERVER_PORT
+        --data-parallel-size 2
+        --data-parallel-size-local 2
+        --tensor-parallel-size 8
+        --seed 1024
+        --quantization ascend
+        --max-num-seqs 16
+        --max-model-len 8192
+        --max-num-batched-tokens 8192
+        --enable-expert-parallel
+        --trust-remote-code
+        --no-enable-prefix-caching
+        --gpu-memory-utilization 0.9
+        --kv-transfer-config
+        '{"kv_connector": "MooncakeConnectorV1",
+        "kv_role": "kv_consumer",
+        "kv_port": "30200",
+        "kv_connector_extra_config": {
+                  "prefill": {
+                          "dp_size": 2,
+                          "tp_size": 8
+                  },
+                  "decode": {
+                          "dp_size": 2,
+                          "tp_size": 8
+                  }
+            }
+        }'
+benchmarks:
+  acc:
+    case_type: accuracy
+    dataset_path: vllm-ascend/gsm8k-lite
+    request_conf: vllm_api_general_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_chat_prompt
+    max_out_len: 4096
+    batch_size: 512
+    baseline: 95
+    threshold: 10
--- a/tests/e2e/weekly/multi_node/internal_dp/config/DeepSeek-V3_2-W8A8-EP_weekly.yaml
+++ b/tests/e2e/weekly/multi_node/internal_dp/config/DeepSeek-V3_2-W8A8-EP_weekly.yaml
@@ -0,0 +1,198 @@
+test_name: "weekly test DeepSeek-V3.2-W8A8-EP disaggregated_prefill"
+model: "vllm-ascend/DeepSeek-V3.2-W8A8"
+num_nodes: 2
+npu_per_node: 16
+env_common: &env_common
+  OMP_PROC_BIND: false
+  OMP_NUM_THREADS: 10
+  HCCL_OP_EXPANSION_MODE: "AIV"
+  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
+  VLLM_USE_V1: "1"
+  HCCL_BUFFSIZE: "360"
+  ASCEND_AGGREGATE_ENABLE: "1"
+  ASCEND_TRANSPORT_PRINT: "1"
+  VLLM_NIXL_ABORT_REQUEST_TIMEOUT: "300000"
+  VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
+  VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
+  VLLM_USE_MODELSCOPE: "true"
+  SERVER_PORT: 8080
+
+disaggregated_prefill:
+  enabled: true
+  prefiller_host_index: [0]
+  decoder_host_index: [1]
+
+deployment:
+  -
+    envs:
+      <<: *env_common
+      VLLM_ASCEND_ENABLE_FLASHCOMM1: 1
+    server_cmd: >
+      vllm serve vllm-ascend/DeepSeek-V3.2-W8A8
+          --host 0.0.0.0
+          --port $SERVER_PORT
+          --data-parallel-size 1
+          --data-parallel-start-rank 0
+          --data-parallel-size-local 1
+          --data-parallel-address $LOCAL_IP
+          --tensor-parallel-size 16
+          --enable-expert-parallel
+          --seed 1024
+          --max-model-len 133000
+          --max-num-batched-tokens 8192
+          --trust-remote-code
+          --max-num-seqs 16
+          --gpu-memory-utilization 0.95
+          --quantization ascend
+          --enforce-eager
+          --no-enable-prefix-caching
+          --tokenizer-mode deepseek_v32 
+          --reasoning-parser deepseek_v3
+          --speculative-config '{"num_speculative_tokens": 2, "method":"deepseek_mtp"}' 
+          --kv-transfer-config
+          '{"kv_connector": "MooncakeConnectorV1",
+          "kv_role": "kv_producer",
+          "kv_port": "30000",
+          "kv_connector_extra_config": {
+                    "prefill": {
+                            "dp_size": 1,
+                            "tp_size": 16
+                    },
+                    "decode": {
+                            "dp_size": 8,
+                            "tp_size": 2
+                    }
+              }
+          }'
+
+  -
+    envs:
+      <<: *env_common
+      OMP_PROC_BIND: false
+      OMP_NUM_THREADS: 10
+      HCCL_OP_EXPANSION_MODE: "AIV"
+      PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
+      VLLM_USE_V1: "1"
+      HCCL_BUFFSIZE: "1100"
+      ASCEND_AGGREGATE_ENABLE: "1"
+      ASCEND_TRANSPORT_PRINT: "1"
+      ACL_OP_INIT_MODE: "1"
+      ASCEND_A3_ENABLE: "1"
+      VLLM_NIXL_ABORT_REQUEST_TIMEOUT: "300000"
+      TASK_QUEUE_ENABLE: "1"
+      VLLM_ASCEND_ENABLE_MLAPO: "1"
+      VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
+      DYNAMIC_EPLB: "true"
+    server_cmd: >
+      vllm serve vllm-ascend/DeepSeek-V3.2-W8A8
+          --host 0.0.0.0
+          --port $SERVER_PORT
+          --data-parallel-size 8
+          --data-parallel-size-local 8
+          --data-parallel-start-rank 0
+          --data-parallel-address $LOCAL_IP
+          --data-parallel-rpc-port 13389
+          --tensor-parallel-size 2
+          --enable-expert-parallel
+          --seed 1024
+          --max-model-len 133000 
+          --max-num-batched-tokens 100
+          --speculative-config '{"num_speculative_tokens": 2, "method":"deepseek_mtp"}'
+          --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY", "cudagraph_capture_sizes":[3,6,9,12,15,18,21,24,27,30,33,36,39,42]}'
+          --additional-config '{"recompute_scheduler_enable": true,"eplb_config": {"dynamic_eplb":true,"expert_heat_collection_interval":2000,"algorithm_execution_interval":50, "eplb_policy_type":3, "num_redundant_experts":0}}'
+          --trust-remote-code
+          --max-num-seqs 14
+          --gpu-memory-utilization 0.92
+          --no-enable-prefix-caching
+          --quantization ascend
+          --tokenizer-mode deepseek_v32
+          --reasoning-parser deepseek_v3
+          --kv-transfer-config
+          '{"kv_connector": "MooncakeConnectorV1",
+          "kv_role": "kv_consumer",
+          "kv_port": "30200",
+          "kv_connector_extra_config": {
+                      "prefill": {
+                              "dp_size": 1,
+                              "tp_size": 16
+                      },
+                      "decode": {
+                              "dp_size": 8,
+                              "tp_size": 2
+                      }
+              }
+          }'
+
+
+benchmarks:
+  perf_16384_1024_num64:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K_prefix0_in16384_bs1200_deepseek
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 64
+    max_out_len: 1024
+    batch_size: 16
+    request_rate: 1
+    baseline: 1
+    threshold: 0.97
+
+  perf_16384_1024_num256:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K_prefix0_in16384_bs1200_deepseek
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 256
+    max_out_len: 1024
+    batch_size: 64
+    request_rate: 1
+    baseline: 1
+    threshold: 0.97
+
+  perf_32768_512_num32:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K_prefix0_in32768_bs1000_deepseek
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 32
+    max_out_len: 512
+    batch_size: 8
+    request_rate: 1
+    baseline: 1
+    threshold: 0.97
+
+  perf_32768_512_num128:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K_prefix0_in32768_bs1000_deepseek
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 128
+    max_out_len: 512
+    batch_size: 32
+    request_rate: 0.7
+    baseline: 1
+    threshold: 0.97
+
+  perf_65536_1024_num4:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K_prefix0_in65536_bs400_deepseek
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 4
+    max_out_len: 1024
+    batch_size: 1
+    request_rate: 0.1
+    baseline: 1
+    threshold: 0.97
+
+  perf_65536_1024_num16:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K_prefix0_in65536_bs400_deepseek
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 16
+    max_out_len: 1024
+    batch_size: 4
+    request_rate: 1
+    baseline: 1
+    threshold: 0.97
--- a/tests/e2e/weekly/multi_node/internal_dp/config/GLM-4.7-W8A8C8-Mooncake-Layerwise.yaml
+++ b/tests/e2e/weekly/multi_node/internal_dp/config/GLM-4.7-W8A8C8-Mooncake-Layerwise.yaml
@@ -0,0 +1,102 @@
+test_name: "test GLM-4.7-W8A8C8 PD separation with mooncake layerwise connector"
+model: "vllm-ascend/GLM-4.7-W8A8C8"
+num_nodes: 2
+npu_per_node: 16
+env_common: &env_common
+  VLLM_USE_MODELSCOPE: true
+  OMP_PROC_BIND: false
+  HCCL_BUFFSIZE: 512
+  HCCL_OP_EXPANSION_MODE: "AIV"
+  SERVER_PORT: 8080
+  DISAGGREGATED_PREFILL_PROXY_SCRIPT: "examples/disaggregated_prefill_v1/load_balance_proxy_layerwise_server_example.py"
+
+disaggregated_prefill:
+  enabled: true
+  prefiller_host_index: [0]
+  decoder_host_index: [1]
+
+deployment:
+  -
+    envs:
+      <<: *env_common
+    server_cmd: >
+        vllm serve "vllm-ascend/GLM-4.7-W8A8C8"
+        --host 0.0.0.0
+        --port $SERVER_PORT
+        --data-parallel-size 2
+        --data-parallel-size-local 2
+        --data-parallel-address $LOCAL_IP
+        --data-parallel-rpc-port 13389
+        --tensor-parallel-size 8
+        --seed 1024
+        --enforce-eager
+        --enable-expert-parallel
+        --max-num-seqs 32
+        --max-model-len 8192
+        --max-num-batched-tokens 8192
+        --trust-remote-code
+        --no-enable-prefix-caching
+        --gpu-memory-utilization 0.9
+        --additional-config '{"enable_shared_expert_dp": true}'
+        --kv-transfer-config
+        '{"kv_connector": "MooncakeLayerwiseConnector",
+        "kv_role": "kv_producer",
+        "kv_port": "30000",
+        "kv_connector_extra_config": {
+                  "prefill": {
+                          "dp_size": 2,
+                          "tp_size": 8
+                  },
+                  "decode": {
+                          "dp_size": 2,
+                          "tp_size": 8
+                  }
+            }
+        }'
+  -
+    envs:
+      <<: *env_common
+    server_cmd: >
+        vllm serve "vllm-ascend/GLM-4.7-W8A8C8"
+        --host 0.0.0.0
+        --port $SERVER_PORT
+        --data-parallel-size 2
+        --data-parallel-size-local 2
+        --data-parallel-address $MASTER_IP
+        --data-parallel-rpc-port 13389
+        --tensor-parallel-size 8
+        --seed 1024
+        --max-num-seqs 16
+        --max-model-len 8192
+        --max-num-batched-tokens 8192
+        --enable-expert-parallel
+        --trust-remote-code
+        --no-enable-prefix-caching
+        --gpu-memory-utilization 0.9
+        --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+        --kv-transfer-config
+        '{"kv_connector": "MooncakeLayerwiseConnector",
+        "kv_role": "kv_consumer",
+        "kv_port": "30200",
+        "kv_connector_extra_config": {
+                  "prefill": {
+                          "dp_size": 2,
+                          "tp_size": 8
+                  },
+                  "decode": {
+                          "dp_size": 2,
+                          "tp_size": 8
+                  }
+            }
+        }'
+
+benchmarks:
+  acc:
+    case_type: accuracy
+    dataset_path: vllm-ascend/gsm8k-lite
+    request_conf: vllm_api_general_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_chat_prompt
+    max_out_len: 4096
+    batch_size: 8
+    baseline: 95
+    threshold: 10
--- a/tests/e2e/weekly/multi_node/internal_dp/config/GLM-5.1-W8A8C8-A3_198k_function.yaml
+++ b/tests/e2e/weekly/multi_node/internal_dp/config/GLM-5.1-W8A8C8-A3_198k_function.yaml
@@ -0,0 +1,104 @@
+#198k长序列只验证功能
+test_name: "multi-node-GLM-5.1-W8A8C8-MTP-A3_198k_function"
+model: "Eco-Tech/GLM-5.1-w8a8c8"
+num_nodes: 2
+npu_per_node: 16
+env_common: &env_common
+  HCCL_OP_EXPANSION_MODE: "AIV"
+  OMP_PROC_BIND: "false"
+  VLLM_USE_MODELSCOPE: "true"
+  OMP_NUM_THREADS: "1"
+  HCCL_BUFFSIZE: "400"
+  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
+  VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
+  VLLM_ASCEND_ENABLE_MLAPO: "1"
+  VLLM_ENGINE_READY_TIMEOUT_S: "3000"
+  VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
+  SERVER_PORT: 8077
+
+deployment:
+  - envs:
+      <<: *env_common
+    server_cmd: >
+      vllm serve Eco-Tech/GLM-5.1-w8a8c8
+      --host 0.0.0.0
+      --port $SERVER_PORT
+      --data-parallel-size 8
+      --data-parallel-size-local 4
+      --data-parallel-start-rank 0
+      --data-parallel-address $LOCAL_IP
+      --enable-expert-parallel
+      --data-parallel-rpc-port 12980
+      --tensor-parallel-size 4
+      --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
+      --seed 1024
+      --tool-call-parser glm47
+      --reasoning-parser glm45
+      --enable-auto-tool-choice
+      --max-num-seqs 6
+      --max-model-len 202752
+      --max-num-batched-tokens 4096
+      --trust-remote-code
+      --gpu-memory-utilization 0.92
+      --quantization ascend
+      --enable-chunked-prefill
+      --enable-prefix-caching
+      --async-scheduling
+      --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+      --additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}' 
+      --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
+  -
+    envs:
+      <<: *env_common
+    server_cmd: >
+      vllm serve Eco-Tech/GLM-5.1-w8a8c8
+      --host 0.0.0.0
+      --port $SERVER_PORT
+      --data-parallel-size 8
+      --data-parallel-size-local 4
+      --data-parallel-start-rank 4
+      --headless
+      --data-parallel-address $MASTER_IP
+      --enable-expert-parallel
+      --data-parallel-rpc-port 12980
+      --tensor-parallel-size 4
+      --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
+      --seed 1024
+      --tool-call-parser glm47
+      --reasoning-parser glm45
+      --enable-auto-tool-choice
+      --max-num-seqs 6
+      --max-model-len 202752
+      --max-num-batched-tokens 4096
+      --trust-remote-code
+      --gpu-memory-utilization 0.9
+      --quantization ascend
+      --enable-chunked-prefill
+      --enable-prefix-caching
+      --async-scheduling
+      --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+      --additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'  
+      --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
+benchmarks:
+  perf:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K-in3500-bs2800
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 8
+    max_out_len: 1500
+    batch_size: 2
+    request_rate: 0
+    baseline: 1
+    threshold: 0.97
+  acc_aime2025:
+    case_type: accuracy
+    dataset_path: vllm-ascend/aime2025
+    request_conf: vllm_api_general_chat
+    dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
+    max_out_len: 129024
+    batch_size: 32
+    baseline: 93.33
+    threshold: 10
+    temperature: 1.0
+    top_p: 0.95
--- a/tests/e2e/weekly/multi_node/internal_dp/config/GLM-5.1-W8A8C8-A3_3.5k_1.5k_0_50.yaml
+++ b/tests/e2e/weekly/multi_node/internal_dp/config/GLM-5.1-W8A8C8-A3_3.5k_1.5k_0_50.yaml
@@ -0,0 +1,105 @@
+test_name: "multi-node-GLM-5.1-W8A8C8-MTP-A3_3.5k_1.5k_0_50"
+model: "Eco-Tech/GLM-5.1-w8a8c8"
+num_nodes: 2
+npu_per_node: 16
+env_common: &env_common
+  HCCL_OP_EXPANSION_MODE: "AIV"
+  OMP_PROC_BIND: "false"
+  VLLM_USE_MODELSCOPE: "true"
+  OMP_NUM_THREADS: "1"
+  HCCL_BUFFSIZE: "400"
+  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
+  VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
+  VLLM_ASCEND_ENABLE_MLAPO: "1"
+  VLLM_ENGINE_READY_TIMEOUT_S: "3000"
+  VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
+  SERVER_PORT: 8077
+
+# TODO: need to identify why TP and mtp+1 divisibility rules break on dual-node case
+
+deployment:
+  - envs:
+      <<: *env_common
+    server_cmd: >
+      vllm serve Eco-Tech/GLM-5.1-w8a8c8
+      --host 0.0.0.0
+      --port $SERVER_PORT
+      --data-parallel-size 8
+      --data-parallel-size-local 4
+      --data-parallel-start-rank 0
+      --data-parallel-address $LOCAL_IP
+      --enable-expert-parallel
+      --data-parallel-rpc-port 12981
+      --tensor-parallel-size 4
+      --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
+      --seed 1024
+      --tool-call-parser glm47
+      --reasoning-parser glm45
+      --enable-auto-tool-choice
+      --max-num-seqs 32
+      --max-model-len 133120
+      --max-num-batched-tokens 4096
+      --trust-remote-code
+      --gpu-memory-utilization 0.92
+      --quantization ascend
+      --enable-chunked-prefill
+      --no-enable-prefix-caching
+      --async-scheduling
+      --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+      --additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
+      --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
+  -
+    envs:
+      <<: *env_common
+    server_cmd: >
+      vllm serve Eco-Tech/GLM-5.1-w8a8c8
+      --host 0.0.0.0
+      --port $SERVER_PORT
+      --data-parallel-size 8
+      --data-parallel-size-local 4
+      --data-parallel-start-rank 4
+      --headless
+      --data-parallel-address $MASTER_IP
+      --enable-expert-parallel
+      --data-parallel-rpc-port 12981
+      --tensor-parallel-size 4
+      --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
+      --seed 1024
+      --tool-call-parser glm47
+      --reasoning-parser glm45
+      --enable-auto-tool-choice
+      --max-num-seqs 32
+      --max-model-len 133120
+      --max-num-batched-tokens 4096
+      --trust-remote-code
+      --gpu-memory-utilization 0.95
+      --quantization ascend
+      --enable-chunked-prefill
+      --no-enable-prefix-caching
+      --async-scheduling
+      --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+      --additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
+      --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
+benchmarks:
+  perf_TPOT50:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K_in3500_bs400
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 256
+    max_out_len: 1500
+    batch_size: 64
+    request_rate: 0
+    baseline: 1267.2573
+    threshold: 0.97
+  perf_TPOT20:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K_in3500_bs400
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 32
+    max_out_len: 1500
+    batch_size: 8
+    request_rate: 0
+    baseline: 345.13
+    threshold: 0.97
--- a/tests/e2e/weekly/multi_node/internal_dp/config/GLM-5.1-W8A8C8-A3_64k_128k_90_50.yaml
+++ b/tests/e2e/weekly/multi_node/internal_dp/config/GLM-5.1-W8A8C8-A3_64k_128k_90_50.yaml
@@ -0,0 +1,115 @@
+test_name: "multi-node-GLM-5.1-W8A8C8-MTP-A3_64k/128k"
+model: "Eco-Tech/GLM-5.1-w8a8c8"
+num_nodes: 2
+npu_per_node: 16
+env_common: &env_common
+  HCCL_OP_EXPANSION_MODE: "AIV"
+  OMP_PROC_BIND: "false"
+  VLLM_USE_MODELSCOPE: "true"
+  OMP_NUM_THREADS: "1"
+  HCCL_BUFFSIZE: "400"
+  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
+  VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
+  VLLM_ASCEND_ENABLE_MLAPO: "1"
+  VLLM_ENGINE_READY_TIMEOUT_S: "3000"
+  VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
+  SERVER_PORT: 8077
+
+# TODO: need to identify why TP and mtp+1 divisibility rules break on dual-node case
+
+deployment:
+  - envs:
+      <<: *env_common
+    server_cmd: >
+      vllm serve Eco-Tech/GLM-5.1-w8a8c8
+      --host 0.0.0.0
+      --port $SERVER_PORT
+      --data-parallel-size 8
+      --data-parallel-size-local 4
+      --data-parallel-start-rank 0
+      --data-parallel-address $LOCAL_IP
+      --enable-expert-parallel
+      --data-parallel-rpc-port 12981
+      --tensor-parallel-size 4
+      --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
+      --seed 1024
+      --tool-call-parser glm47
+      --reasoning-parser glm45
+      --enable-auto-tool-choice
+      --max-num-seqs 6
+      --max-model-len 133120
+      --max-num-batched-tokens 4096
+      --trust-remote-code
+      --gpu-memory-utilization 0.92
+      --quantization ascend
+      --enable-chunked-prefill
+      --enable-prefix-caching
+      --async-scheduling
+      --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+      --additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
+      --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
+  - envs:
+      <<: *env_common
+    server_cmd: >
+      vllm serve Eco-Tech/GLM-5.1-w8a8c8
+      --host 0.0.0.0
+      --port $SERVER_PORT
+      --data-parallel-size 8
+      --data-parallel-size-local 4
+      --data-parallel-start-rank 4
+      --headless
+      --data-parallel-address $MASTER_IP
+      --enable-expert-parallel
+      --data-parallel-rpc-port 12981
+      --tensor-parallel-size 4
+      --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
+      --seed 1024
+      --tool-call-parser glm47
+      --reasoning-parser glm45
+      --enable-auto-tool-choice
+      --max-num-seqs 6
+      --max-model-len 133120
+      --max-num-batched-tokens 4096
+      --trust-remote-code
+      --gpu-memory-utilization 0.9
+      --quantization ascend
+      --enable-chunked-prefill
+      --enable-prefix-caching
+      --async-scheduling
+      --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+      --additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
+      --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
+benchmarks:
+  perf_64k:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K-in65536-bs1500-prefix90-glm51
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 128
+    max_out_len: 1024
+    batch_size: 32
+    request_rate: 0
+    baseline: 539.11
+    threshold: 0.97
+  perf_128k_warmup:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K-in131072-bs500-prefix90-glm51
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 1
+    max_out_len: 1
+    batch_size: 1
+    request_rate: 0
+    baseline: 0
+    threshold: 0.97
+  perf_128k:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K-in131072-bs500-prefix90-glm51
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 256
+    max_out_len: 1024
+    batch_size: 64
+    request_rate: 0
+    baseline: 290.8308
+    threshold: 0.97
--- a/tests/e2e/weekly/multi_node/internal_dp/config/GLM-5.2-W4A8C8-A3_198k_function.yaml
+++ b/tests/e2e/weekly/multi_node/internal_dp/config/GLM-5.2-W4A8C8-A3_198k_function.yaml
@@ -0,0 +1,104 @@
+#198k长序列只验证功能
+test_name: "multi-node-GLM-5.2-W4A8C8-MTP-A3_198k_function"
+model: "Eco-Tech/GLM-5.2-w4a8c8"
+num_nodes: 2
+npu_per_node: 16
+env_common: &env_common
+  HCCL_OP_EXPANSION_MODE: "AIV"
+  OMP_PROC_BIND: "false"
+  VLLM_USE_MODELSCOPE: "true"
+  OMP_NUM_THREADS: "1"
+  HCCL_BUFFSIZE: "400"
+  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
+  VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
+  VLLM_ASCEND_ENABLE_MLAPO: "1"
+  VLLM_ENGINE_READY_TIMEOUT_S: "3000"
+  VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
+  SERVER_PORT: 8077
+
+deployment:
+  - envs:
+      <<: *env_common
+    server_cmd: >
+      vllm serve Eco-Tech/GLM-5.2-w4a8c8
+      --host 0.0.0.0
+      --port $SERVER_PORT
+      --data-parallel-size 8
+      --data-parallel-size-local 4
+      --data-parallel-start-rank 0
+      --data-parallel-address $LOCAL_IP
+      --enable-expert-parallel
+      --data-parallel-rpc-port 12980
+      --tensor-parallel-size 4
+      --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
+      --seed 1024
+      --tool-call-parser glm47
+      --reasoning-parser glm45
+      --enable-auto-tool-choice
+      --max-num-seqs 6
+      --max-model-len 202752
+      --max-num-batched-tokens 4096
+      --trust-remote-code
+      --gpu-memory-utilization 0.92
+      --quantization ascend
+      --enable-chunked-prefill
+      --enable-prefix-caching
+      --async-scheduling
+      --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+      --additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}' 
+      --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
+  -
+    envs:
+      <<: *env_common
+    server_cmd: >
+      vllm serve Eco-Tech/GLM-5.2-w4a8c8
+      --host 0.0.0.0
+      --port $SERVER_PORT
+      --data-parallel-size 8
+      --data-parallel-size-local 4
+      --data-parallel-start-rank 4
+      --headless
+      --data-parallel-address $MASTER_IP
+      --enable-expert-parallel
+      --data-parallel-rpc-port 12980
+      --tensor-parallel-size 4
+      --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
+      --seed 1024
+      --tool-call-parser glm47
+      --reasoning-parser glm45
+      --enable-auto-tool-choice
+      --max-num-seqs 6
+      --max-model-len 202752
+      --max-num-batched-tokens 4096
+      --trust-remote-code
+      --gpu-memory-utilization 0.9
+      --quantization ascend
+      --enable-chunked-prefill
+      --enable-prefix-caching
+      --async-scheduling
+      --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+      --additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": true, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'  
+      --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
+benchmarks:
+  perf:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K-in3500-bs2800
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 8
+    max_out_len: 1500
+    batch_size: 2
+    request_rate: 0
+    baseline: 1
+    threshold: 0.97
+  acc_aime2025:
+    case_type: accuracy
+    dataset_path: vllm-ascend/aime2025
+    request_conf: vllm_api_general_chat
+    dataset_conf: aime2025/aime2025_gen_0_shot_chat_prompt
+    max_out_len: 129024
+    batch_size: 32
+    baseline: 93.33
+    threshold: 10
+    temperature: 1.0
+    top_p: 0.95
--- a/tests/e2e/weekly/multi_node/internal_dp/config/GLM-5.2-W4A8C8-A3_3.5k_1.5k_0_50.yaml
+++ b/tests/e2e/weekly/multi_node/internal_dp/config/GLM-5.2-W4A8C8-A3_3.5k_1.5k_0_50.yaml
@@ -0,0 +1,105 @@
+test_name: "multi-node-GLM-5.2-W4A8C8-MTP-A3_3.5k_1.5k_0_50"
+model: "Eco-Tech/GLM-5.2-w4a8c8"
+num_nodes: 2
+npu_per_node: 16
+env_common: &env_common
+  HCCL_OP_EXPANSION_MODE: "AIV"
+  OMP_PROC_BIND: "false"
+  VLLM_USE_MODELSCOPE: "true"
+  OMP_NUM_THREADS: "1"
+  HCCL_BUFFSIZE: "400"
+  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
+  VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
+  VLLM_ASCEND_ENABLE_MLAPO: "1"
+  VLLM_ENGINE_READY_TIMEOUT_S: "3000"
+  VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
+  SERVER_PORT: 8077
+
+# TODO: need to identify why TP and mtp+1 divisibility rules break on dual-node case
+
+deployment:
+  - envs:
+      <<: *env_common
+    server_cmd: >
+      vllm serve Eco-Tech/GLM-5.2-w4a8c8
+      --host 0.0.0.0
+      --port $SERVER_PORT
+      --data-parallel-size 8
+      --data-parallel-size-local 4
+      --data-parallel-start-rank 0
+      --data-parallel-address $LOCAL_IP
+      --enable-expert-parallel
+      --data-parallel-rpc-port 12981
+      --tensor-parallel-size 4
+      --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
+      --seed 1024
+      --tool-call-parser glm47
+      --reasoning-parser glm45
+      --enable-auto-tool-choice
+      --max-num-seqs 32
+      --max-model-len 133120
+      --max-num-batched-tokens 4096
+      --trust-remote-code
+      --gpu-memory-utilization 0.92
+      --quantization ascend
+      --enable-chunked-prefill
+      --no-enable-prefix-caching
+      --async-scheduling
+      --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+      --additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
+      --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
+  -
+    envs:
+      <<: *env_common
+    server_cmd: >
+      vllm serve Eco-Tech/GLM-5.2-w4a8c8
+      --host 0.0.0.0
+      --port $SERVER_PORT
+      --data-parallel-size 8
+      --data-parallel-size-local 4
+      --data-parallel-start-rank 4
+      --headless
+      --data-parallel-address $MASTER_IP
+      --enable-expert-parallel
+      --data-parallel-rpc-port 12981
+      --tensor-parallel-size 4
+      --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
+      --seed 1024
+      --tool-call-parser glm47
+      --reasoning-parser glm45
+      --enable-auto-tool-choice
+      --max-num-seqs 32
+      --max-model-len 133120
+      --max-num-batched-tokens 4096
+      --trust-remote-code
+      --gpu-memory-utilization 0.95
+      --quantization ascend
+      --enable-chunked-prefill
+      --no-enable-prefix-caching
+      --async-scheduling
+      --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+      --additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
+      --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
+benchmarks:
+  perf_TPOT50:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K_in3500_bs400
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 256
+    max_out_len: 1500
+    batch_size: 64
+    request_rate: 0
+    baseline: 1267.2573
+    threshold: 0.97
+  perf_TPOT20:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K_in3500_bs400
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 32
+    max_out_len: 1500
+    batch_size: 8
+    request_rate: 0
+    baseline: 345.13
+    threshold: 0.97
--- a/tests/e2e/weekly/multi_node/internal_dp/config/GLM-5.2-W4A8C8-A3_64k_128k_90_50.yaml
+++ b/tests/e2e/weekly/multi_node/internal_dp/config/GLM-5.2-W4A8C8-A3_64k_128k_90_50.yaml
@@ -0,0 +1,105 @@
+test_name: "multi-node-GLM-5.2-W4A8C8-MTP-A3_64k/128k"
+model: "Eco-Tech/GLM-5.2-w4a8c8"
+num_nodes: 2
+npu_per_node: 16
+env_common: &env_common
+  HCCL_OP_EXPANSION_MODE: "AIV"
+  OMP_PROC_BIND: "false"
+  VLLM_USE_MODELSCOPE: "true"
+  OMP_NUM_THREADS: "1"
+  HCCL_BUFFSIZE: "400"
+  PYTORCH_NPU_ALLOC_CONF: "expandable_segments:True"
+  VLLM_ASCEND_ENABLE_FLASHCOMM1: "1"
+  VLLM_ASCEND_ENABLE_MLAPO: "1"
+  VLLM_ENGINE_READY_TIMEOUT_S: "3000"
+  VLLM_ASCEND_ENABLE_FUSED_MC2: "1"
+  SERVER_PORT: 8077
+
+# TODO: need to identify why TP and mtp+1 divisibility rules break on dual-node case
+
+deployment:
+  - envs:
+      <<: *env_common
+    server_cmd: >
+      vllm serve Eco-Tech/GLM-5.2-w4a8c8
+      --host 0.0.0.0
+      --port $SERVER_PORT
+      --data-parallel-size 8
+      --data-parallel-size-local 4
+      --data-parallel-start-rank 0
+      --data-parallel-address $LOCAL_IP
+      --enable-expert-parallel
+      --data-parallel-rpc-port 12981
+      --tensor-parallel-size 4
+      --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
+      --seed 1024
+      --tool-call-parser glm47
+      --reasoning-parser glm45
+      --enable-auto-tool-choice
+      --max-num-seqs 6
+      --max-model-len 133120
+      --max-num-batched-tokens 4096
+      --trust-remote-code
+      --gpu-memory-utilization 0.92
+      --quantization ascend
+      --enable-chunked-prefill
+      --enable-prefix-caching
+      --async-scheduling
+      --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+      --additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
+      --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
+  -
+    envs:
+      <<: *env_common
+    server_cmd: >
+      vllm serve Eco-Tech/GLM-5.2-w4a8c8
+      --host 0.0.0.0
+      --port $SERVER_PORT
+      --data-parallel-size 8
+      --data-parallel-size-local 4
+      --data-parallel-start-rank 4
+      --headless
+      --data-parallel-address $MASTER_IP
+      --enable-expert-parallel
+      --data-parallel-rpc-port 12981
+      --tensor-parallel-size 4
+      --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
+      --seed 1024
+      --tool-call-parser glm47
+      --reasoning-parser glm45
+      --enable-auto-tool-choice
+      --max-num-seqs 6
+      --max-model-len 133120
+      --max-num-batched-tokens 4096
+      --trust-remote-code
+      --gpu-memory-utilization 0.9
+      --quantization ascend
+      --enable-chunked-prefill
+      --enable-prefix-caching
+      --async-scheduling
+      --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+      --additional-config '{"enable_dsa_cp": true, "enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true, "enable_balance_scheduling": true, "fuse_muls_add": true, "multistream_overlap_shared_expert": true}'
+      --speculative-config '{"num_speculative_tokens": 3, "method": "deepseek_mtp","enforce_eager":true}'
+benchmarks:
+  perf_64k:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K-in65536-bs1500-prefix90-glm51
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 128
+    max_out_len: 1024
+    batch_size: 32
+    request_rate: 0
+    baseline: 539.11
+    threshold: 0.97
+  perf_128k:
+    case_type: performance
+    dataset_path: vllm-ascend/GSM8K-in131072-bs500-prefix90-glm51
+    request_conf: vllm_api_stream_chat
+    dataset_conf: gsm8k/gsm8k_gen_0_shot_cot_str_perf
+    num_prompts: 256
+    max_out_len: 1024
+    batch_size: 64
+    request_rate: 0
+    baseline: 796.35
+    threshold: 0.97