Add study tune loop and smoke configs

configs/examples/dash0_smoke_proposals/aggressive.json

@@ -0,0 +1,19 @@
+{
+  "observation": "Push batching further after validating the balanced layout.",
+  "diagnosis": "If TTFT remains under control, a larger admission window should maximize achieved request rate.",
+  "config_patch": {
+    "env_patch": {},
+    "flag_patch": {
+      "tensor-parallel-size": 4,
+      "max-num-seqs": 24,
+      "max-num-batched-tokens": 98304,
+      "gpu-memory-utilization": 0.92,
+      "block-size": 64
+    }
+  },
+  "expected_effects": [
+    "Highest throughput among the smoke candidates if memory is sufficient",
+    "More pressure on TTFT, so the binary search should locate the safe threshold"
+  ],
+  "why_not_previous_failures": "Keeps tp=4 and block-size stable while only expanding batching and memory utilization."
+}

configs/examples/dash0_smoke_proposals/balanced.json

@@ -0,0 +1,19 @@
+{
+  "observation": "Increase batching once tp=4 is fixed.",
+  "diagnosis": "Throughput should improve if the engine can admit more concurrent prefills without violating TTFT.",
+  "config_patch": {
+    "env_patch": {},
+    "flag_patch": {
+      "tensor-parallel-size": 4,
+      "max-num-seqs": 16,
+      "max-num-batched-tokens": 65536,
+      "gpu-memory-utilization": 0.9,
+      "block-size": 64
+    }
+  },
+  "expected_effects": [
+    "Higher feasible sampling_u than the conservative baseline",
+    "Better token throughput if memory headroom is sufficient"
+  ],
+  "why_not_previous_failures": "Raises batching in a controlled step instead of jumping directly to the most aggressive setting."
+}

configs/examples/dash0_smoke_proposals/conservative.json

@@ -0,0 +1,19 @@
+{
+  "observation": "Start from a safe tp=4 layout and conservative batching.",
+  "diagnosis": "The first pass should verify multi-GPU launch and avoid queueing collapse from over-batching.",
+  "config_patch": {
+    "env_patch": {},
+    "flag_patch": {
+      "tensor-parallel-size": 4,
+      "max-num-seqs": 8,
+      "max-num-batched-tokens": 32768,
+      "gpu-memory-utilization": 0.85,
+      "block-size": 64
+    }
+  },
+  "expected_effects": [
+    "Stable startup on 4x H20",
+    "Low risk of OOM during the first binary-search probes"
+  ],
+  "why_not_previous_failures": "This is the initial baseline proposal."
+}

configs/examples/dash0_smoke_study.json

@@ -0,0 +1,83 @@
+{
+  "study_id": "dash0-qwen30b-chat-smoke",
+  "hardware": {
+    "gpu_count": 4,
+    "gpu_model": "H20",
+    "host_candidates": ["dash0"]
+  },
+  "model": {
+    "model_id": "qwen3-30b-a3b",
+    "served_model_name": "qwen3-30b-smoke"
+  },
+  "engine": {
+    "engine_name": "vllm",
+    "engine_version": "0.13.0rc2.dev2111+gb44b43f43.d20260309",
+    "exec_path": "/usr/local/bin/vllm",
+    "cwd": "/home/admin/cpfs/wjh/aituner/aituner",
+    "host": "127.0.0.1",
+    "port": 18080,
+    "healthcheck_path": "/v1/models",
+    "ready_timeout_s": 900,
+    "request_timeout_s": 900,
+    "launch_args": [
+      "serve",
+      "/home/admin/resource/model/464482ce.qwen3-30b-a3b/1m-instruct-0726-fp4"
+    ],
+    "base_envs": {
+      "CUDA_VISIBLE_DEVICES": "0,1,2,3"
+    },
+    "base_flags": {
+      "host": "127.0.0.1",
+      "port": 18080,
+      "served-model-name": "qwen3-30b-smoke",
+      "max-model-len": 65536,
+      "disable-log-requests": true,
+      "trust-remote-code": true
+    },
+    "tunable_envs": [
+      "VLLM_ATTENTION_BACKEND"
+    ],
+    "tunable_flags": [
+      "tensor-parallel-size",
+      "max-num-seqs",
+      "max-num-batched-tokens",
+      "gpu-memory-utilization",
+      "block-size"
+    ],
+    "python_executable": "python3"
+  },
+  "trace": {
+    "windows_path": "/home/admin/cpfs/wjh/aituner/aituner/trace_windows/windows.json",
+    "window_id": "chat_w20260311_1000",
+    "u_field": "sampling_u",
+    "timestamp_field": "timestamp",
+    "max_concurrency": 2,
+    "max_requests_per_probe": 24
+  },
+  "slo": {
+    "target_pass_rate": 0.95,
+    "ttft_rule": {
+      "kind": "step_ms",
+      "buckets": [
+        {"max_input_tokens": 4096, "threshold_ms": 15000},
+        {"max_input_tokens": 16384, "threshold_ms": 30000},
+        {"threshold_ms": 45000}
+      ]
+    },
+    "tpot_rule": {
+      "kind": "fixed_ms",
+      "threshold_ms": 1500
+    }
+  },
+  "search": {
+    "low": 0.0,
+    "high": 1.0,
+    "tolerance": 0.1,
+    "max_probes": 4,
+    "sample_seed": 20260325
+  },
+  "llm": {
+    "system_prompt": "Propose a single engine config patch that increases the maximum feasible sampling_u under the SLO target.",
+    "max_history_trials": 8
+  }
+}