{ "study_id": "dash0-qwen235b-decode-thinking-run2-tpot40", "hardware": { "gpu_count": 8, "gpu_model": "H20", "host_candidates": [ "dash0" ] }, "model": { "model_id": "qwen3-235b-a22b-256k-0717-internal", "served_model_name": "qwen3-235b-decode-aituner" }, "engine": { "engine_name": "vllm", "engine_version": "internal-on-dash0", "exec_path": "/usr/local/bin/vllm", "cwd": "/home/admin/cpfs/wjh/aituner/aituner", "host": "127.0.0.1", "port": 18120, "healthcheck_path": "/v1/models", "ready_timeout_s": 1800, "request_timeout_s": 1800, "launch_args": [ "serve", "/home/admin/resource/model/464482ce.qwen3-235b-a22b/256k-0717" ], "base_envs": { "CUDA_VISIBLE_DEVICES": "0,1,2,3,4,5,6,7", "DS_LLM_MULTI_ENGINE_NUM": "2", "DS_GPU_NUM": "8", "DS_LLM_GRACEFUL_SHUTDOWN_WAIT_SECONDS": "600", "DASHGEN_DEPLOYMENT_ROLE": "decode", "DS_LLM_IGNORE_WARMUP": "1", "DS_MODEL_PRELOAD_TO_SHM": "1", "DS_LLM_PD_DECODE_FIRST_TIMEOUT_TIME": "120", "DS_LLM_SERVER_MAX_CONCURRENCY": "256", "AQUILA_RPC_ENABLE_WSPP": "1", "AQUILA_HEALTHY_PROCESS_TIME_AVG_THRESHOLD": "7200000", "VLLM_FUSE_QKNORM_AND_ROPE": "1", "VLLM_FUSE_QKNORM_ROPE_AND_KVCACHE_WRITE": "1", "VLLM_FP8_USE_BLADNN": "0", "VLLM_MOE_USE_BLADNN": "0", "VLLM_USE_V1": "1", "VLLM_ENABLE_TORCH_COMPILE": "1", "VLLM_ATTENTION_BACKEND": "FLASH_ATTN", "VLLM_QUANTIZATION_LAYER_WISE": "1", "VLLM_MOE_USE_DEEPEP": "1", "VLLM_ENABLE_TBO_OPT": "0", "VLLM_MOE_BALANCED_GATING": "0", "VLLM_MOE_RANDOM_GATING": "0", "VLLM_FUSED_MOE_CHUNK_SIZE": "4096", "VLLM_DP_META_USE_CPU_GROUP": "0", "VLLM_MLA_FP8_ATTENTION": "0", "VLLM_MOE_EXPERTS_OVERLAP": "1", "VLLM_USE_FLASHINFER_SAMPLER": "0", "VLLM_DP_MASTER_PORT": "9524", "VLLM_RESPONSE_TIMEOUT": "120", "VLLM_PD_TRY_CONNECT_TIMEOUT_SECONDS": "150", "VLLM_KVT_MAX_DELAY_MS": "2000", "VLLM_DEEP_GEMM_WARMUP": "skip", "TORCH_CUDA_ARCH_LIST": "9.0+PTX", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "ENABLE_SWAPAB": "1", "ACCL_WRITEBATCH_OPT": "2", "ACCL_IBV_MTU": "9000", "ACCL_TX_DEPTH": "1024", "ACCL_RETRANSMIT_TIMEOUT": "17", "ACCL_C4_STATS_MODE": "CONN", "ACCL_IB_SPLIT_DATA_NUM": "4", "ACCL_IB_QPS_LOAD_BALANCE": "1", "ACCL_IB_GID_INDEX_FIX": "1", "ACCL_LOG_TIME": "1", "NCCL_NVLS_ENABLE": "0", "NCCL_CUMEM_ENABLE": "0", "ACCL_NORMAL_MODE": "ibrc", "ACCL_LOAD_BALANCE": "1", "ACCL_TOPO_FIX": "1", "ACCL_LOW_LATENCY_COMBINE_USE_FP8": "1", "ACCL_LOW_LATENCY_BUFFER_FP8_OPT": "1", "ACCL_LOW_LATENCY_BUFFER_USE_SINGLE": "1", "ACCL_DISPATCH_NUM_WARP_GROUPS": "4", "ACCL_COMBINE_NUM_WARP_GROUPS": "4", "ACCL_LOW_LATENCY_OPTIMIZE": "3", "BLLM_KVTRANS_RDMA_SP": "2", "NCCL_IB_TC": "136", "NCCL_IB_SL": "5", "NCCL_IB_GID_INDEX": "3", "NCCL_SOCKET_IFNAME": "eth1", "NCCL_SOCKET_FAMILY": "AF_INET", "NCCL_DEBUG": "WARN", "NCCL_DEBUG_SUBSYS": "TUNING", "NCCL_IB_HCA": "mlx5", "NCCL_IB_TIMEOUT": "22", "NCCL_IB_QPS_PER_CONNECTION": "4", "NCCL_MIN_NCHANNELS": "2", "NCCL_NET_PLUGIN": "none", "NVSHMEM_ENABLE_NIC_PE_MAPPING": "1", "NVSHMEM_HCA_PE_MAPPING": "mlx5_0:1:2,mlx5_1:1:2,mlx5_2:1:2,mlx5_3:1:2", "NVSHMEM_IBRC_ROCE_LAG_PORT_SELECTION": "3", "NVSHMEM_IB_GID_INDEX": "3", "NVSHMEM_IB_ENABLE_IBGDA": "1", "NVSHMEM_IB_TRAFFIC_CLASS": "16", "NVSHMEM_BOOTSTRAP_UID_SOCK_IFNAME": "eth1", "NVSHMEM_IBGDA_NUM_RC_PER_PE": "4" }, "base_flags": { "host": "127.0.0.1", "port": 18120, "served-model-name": "qwen3-235b-decode-aituner", "gpu-memory-utilization": 0.75, "max-model-len": 262144, "enable-chunked-prefill": true, "speculative-config": "{\"method\":\"eagle3\",\"num_speculative_tokens\":2,\"hf_overrides\":{\"rope_scaling\":{\"type\":\"yarn\",\"factor\":128,\"original_max_position_embeddings\":2048,\"semi_dynamic\":false,\"dynamic\":true},\"num_experts\":0},\"model\":\"/home/admin/resource/model/464482ce.qwen3-235b-a22b/0717-eagle-0820\"}", "enable-prefix-caching": true, "max-num-batched-tokens": 1024, "enable-expert-parallel": true, "block-size": 64, "max-num-seqs": 192, "disable-custom-all-reduce": true, "quantization": "fp8", "expert-parallel-size": 8, "data-parallel-size": 2, "tensor-parallel-size": 4, "cuda-graph-sizes": [ 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 40, 48, 64, 128, 192 ], "compilation-config": "{\"cudagraph_mode\":\"FULL_DECODE_ONLY\",\"use_inductor\":true}", "hf-overrides": "{\"architectures\":[\"Qwen3MoeForCausalLM\"],\"model_type\":\"qwen3_moe\"}", "kv-cache-dtype": "fp8", "disable-hybrid-kv-cache-manager": true, "kv-transfer-config": "{\"kv_connector\":\"DecodeBenchConnector\",\"kv_role\":\"kv_both\"}", "disable-log-requests": true }, "tunable_envs": [ "VLLM_ENABLE_TORCH_COMPILE", "VLLM_ENABLE_TBO_OPT", "VLLM_USE_FLASHINFER_SAMPLER", "CUDA_DEVICE_MAX_CONNECTIONS" ], "tunable_flags": [ "tensor-parallel-size", "data-parallel-size", "expert-parallel-size", "gpu-memory-utilization", "max-num-batched-tokens", "max-num-seqs", "block-size" ], "topology_constraints": { "require_tp_dp_product_equals_gpu_count": true, "require_ep_size_leq_tp_dp_product": true, "require_ep_size_divides_tp_dp_product": true, "require_enable_expert_parallel_when_ep_gt_one": true, "validate_cuda_graph_sizes_divisible_by_tp_when_tp_ep_reduce_scatter": true, "allowed_tensor_parallel_sizes": [1, 2, 4, 8], "allowed_data_parallel_sizes": [1, 2, 4, 8], "allowed_expert_parallel_sizes": [1, 2, 4, 8] }, "python_executable": "python3" }, "trace": { "windows_path": "/home/admin/cpfs/wjh/aituner/aituner/trace_windows/windows.json", "window_id": "thinking_w20260327_1000", "request_mode": "decode_only", "u_field": "sampling_u", "timestamp_field": "timestamp", "max_concurrency": 128, "replay_time_scale": 1.0, "early_stop_max_lag_s": 180.0, "early_stop_max_elapsed_s": 1200.0 }, "slo": { "target_pass_rate": 0.95, "tpot_rule": { "kind": "fixed_ms", "threshold_ms": 40 } }, "search": { "low": 0.0, "high": 0.125, "tolerance": 0.001, "max_probes": 6, "sample_seed": 20260325 }, "llm": { "system_prompt": "You are tuning a decode-only vLLM serving stack. TTFT is not an enforced objective unless the study says so. Propose one launch-safe config patch that increases the maximum feasible sampling_u while keeping TPOT pass rate at or above target.", "max_history_trials": 8, "endpoint": { "provider": "codex", "model": "gpt-5.4", "stream": true, "api_key_env": "OPENAI_API_KEY", "timeout_s": 240 } } }