Publish selected TP4 reference, full speed evidence and matched FP8/NVFP4 cache KLD (part 4)

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/decode8k-c8-command.json

@@ -0,0 +1,29 @@
+[
+  "/usr/bin/python3",
+  "<campaign>/fixed-workload-candidate-comparison-01/bench_fixed.py",
+  "--host",
+  "127.0.0.1",
+  "--port",
+  "8001",
+  "--model",
+  "glm53-flash-trellismx-p8-k45",
+  "--duration",
+  "20",
+  "--max-tokens",
+  "8192",
+  "--token-targeting",
+  "exact",
+  "--display-mode",
+  "plain",
+  "--output",
+  "<campaign>/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/decode8k-c8.json",
+  "--contexts",
+  "8k",
+  "--concurrency",
+  "8",
+  "--skip-prefill",
+  "--cell-warmup-timeout-seconds",
+  "180",
+  "--temperature",
+  "0"
+]

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/decode8k-c8-cooldown.json

@@ -0,0 +1,84 @@
+{
+  "status": "passed",
+  "minimum_seconds": 90,
+  "temperature_ceiling_c": 55,
+  "stable_seconds": 30,
+  "samples": [
+    {
+      "elapsed_seconds": 2.8102658689022064e-07,
+      "hardware_csv": "0, 65, 2310, 88.71, 300.00\n1, 45, 2857, 134.00, 300.00\n2, 54, 2317, 78.59, 300.00\n3, 49, 2850, 133.57, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 5.106069390894845,
+      "hardware_csv": "0, 63, 2272, 84.57, 300.00\n1, 43, 2610, 87.93, 300.00\n2, 52, 2272, 76.64, 300.00\n3, 47, 2610, 99.85, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 10.208308120956644,
+      "hardware_csv": "0, 61, 2272, 81.58, 300.00\n1, 42, 2610, 87.21, 300.00\n2, 51, 2272, 75.87, 300.00\n3, 46, 2610, 99.34, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 15.314181811874732,
+      "hardware_csv": "0, 59, 427, 66.26, 300.00\n1, 41, 420, 64.92, 300.00\n2, 49, 2272, 55.00, 300.00\n3, 45, 900, 92.50, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 20.4524090338964,
+      "hardware_csv": "0, 57, 180, 18.06, 300.00\n1, 40, 180, 16.95, 300.00\n2, 47, 180, 10.60, 300.00\n3, 44, 180, 29.55, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 25.593484913930297,
+      "hardware_csv": "0, 56, 180, 16.21, 300.00\n1, 39, 180, 15.71, 300.00\n2, 46, 180, 10.84, 300.00\n3, 43, 180, 26.59, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 30.7240683529526,
+      "hardware_csv": "0, 55, 180, 14.93, 300.00\n1, 39, 180, 15.59, 300.00\n2, 45, 180, 10.58, 300.00\n3, 42, 180, 26.41, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 35.865257401019335,
+      "hardware_csv": "0, 54, 180, 13.90, 300.00\n1, 38, 180, 15.76, 300.00\n2, 44, 180, 9.45, 300.00\n3, 42, 180, 23.11, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 41.00044325296767,
+      "hardware_csv": "0, 53, 180, 12.61, 300.00\n1, 38, 180, 15.53, 300.00\n2, 43, 180, 9.44, 300.00\n3, 41, 180, 22.46, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 46.131547750905156,
+      "hardware_csv": "0, 52, 180, 12.60, 300.00\n1, 37, 180, 15.62, 300.00\n2, 42, 180, 9.24, 300.00\n3, 41, 180, 21.49, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 51.264180578058586,
+      "hardware_csv": "0, 52, 180, 13.30, 300.00\n1, 37, 180, 15.51, 300.00\n2, 42, 180, 8.62, 300.00\n3, 40, 180, 22.92, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 56.39971108804457,
+      "hardware_csv": "0, 51, 180, 12.02, 300.00\n1, 36, 180, 15.52, 300.00\n2, 41, 180, 8.48, 300.00\n3, 40, 180, 21.36, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 61.53315005893819,
+      "hardware_csv": "0, 50, 180, 12.99, 300.00\n1, 36, 180, 15.53, 300.00\n2, 40, 180, 7.93, 300.00\n3, 39, 180, 23.29, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 66.6683305578772,
+      "hardware_csv": "0, 50, 180, 12.86, 300.00\n1, 35, 172, 15.37, 300.00\n2, 40, 180, 7.92, 300.00\n3, 39, 180, 21.10, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 71.8079784088768,
+      "hardware_csv": "0, 49, 180, 11.80, 300.00\n1, 35, 180, 15.54, 300.00\n2, 39, 180, 7.79, 300.00\n3, 38, 180, 22.64, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 76.94549162103795,
+      "hardware_csv": "0, 48, 180, 11.85, 300.00\n1, 35, 180, 15.50, 300.00\n2, 39, 180, 7.67, 300.00\n3, 38, 180, 21.60, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 82.0760404379107,
+      "hardware_csv": "0, 48, 180, 10.76, 300.00\n1, 34, 180, 15.34, 300.00\n2, 38, 180, 7.43, 300.00\n3, 38, 180, 20.24, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 87.21064653596841,
+      "hardware_csv": "0, 47, 180, 12.34, 300.00\n1, 34, 180, 15.13, 300.00\n2, 38, 180, 7.05, 300.00\n3, 37, 180, 22.81, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 92.34947287291288,
+      "hardware_csv": "0, 47, 180, 10.47, 300.00\n1, 34, 180, 15.13, 300.00\n2, 37, 180, 7.10, 300.00\n3, 37, 180, 21.71, 300.00\n"
+    }
+  ]
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/decode8k-c8-fixed-workload.json

@@ -0,0 +1,5 @@
+{
+  "prefix": "tmxrepeatabc",
+  "temperature": 0,
+  "scope": "greedy fixed-prompt diagnostic"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/decode8k-c8-receipt.json

@@ -0,0 +1,5 @@
+{
+  "exit_code": 0,
+  "result_exists": true,
+  "sha256": "e149812699d13d9b17389fcc45cf02df4e03b4c19b234c8f4e6b63b285ae8853"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/decode8k-c8.json

@@ -0,0 +1,429 @@
+{
+  "metadata": {
+    "version": "0.4.29",
+    "engine": "vllm",
+    "model": "glm53-flash-trellismx-p8-k45",
+    "server": "127.0.0.1:8001",
+    "timestamp": "2026-09-09T14:36:35.049940",
+    "decode_mode": "duration",
+    "primary_decode_layer": "sustained_decode",
+    "duration_per_test": 20.0,
+    "request_count": 0,
+    "warmup_request_count": 0,
+    "run_burst": false,
+    "prefill_mode": "skipped",
+    "standalone_prefill": false,
+    "prefill_only": false,
+    "skip_prefill": true,
+    "burst_e2e_status": "not_run_use_--run-burst",
+    "burst_request_count": 0,
+    "burst_warmup_request_count": 0,
+    "burst_requests_per_concurrency": 5,
+    "decode_warmup_seconds": 3.0,
+    "decode_warmup_context": 8192,
+    "decode_warmup_concurrency": 1,
+    "cell_warmup_timeout_seconds": 180.0,
+    "cell_warmup_timeout_policy": "<=32k:60s,64k:120s,>=128k:180s when override is 0",
+    "show_capacity_limited_values": false,
+    "max_tokens": 8192,
+    "temperature": 0.0,
+    "ignore_eos": true,
+    "max_total_tokens": 29532160,
+    "dcp_size": 0,
+    "metrics_available": true,
+    "metrics_warning": "",
+    "concurrency_levels": [
+      8
+    ],
+    "context_lengths": [
+      8192
+    ],
+    "startup_diagnostics_available": true,
+    "nvidia_p2p_override_effective": true,
+    "p2pmark_status": "not_run",
+    "amd_fabric_status": "not_run"
+  },
+  "startup_diagnostics": {
+    "version": "0.4.29",
+    "server_url": "http://127.0.0.1:8001",
+    "hostname": "<host>",
+    "uname": "Linux <host> 6.18.7-76061807-generic #202601231045~1769703228~24.04~cb87b5b SMP PREEMPT_DYNAMIC Thu J x86_64 x86_64 x86_64 GNU/Linux",
+    "env": {},
+    "args": {
+      "concurrency": "8",
+      "contexts": "8k",
+      "max_tokens": 8192,
+      "duration": 20.0,
+      "request_count": 0,
+      "run_burst": false,
+      "standalone_prefill": false,
+      "prefill_only": false,
+      "skip_prefill": true,
+      "prefill_contexts": "8k,64k,128k",
+      "prefill_metric": "client",
+      "dcp_size": 0,
+      "kv_budget": 0
+    },
+    "nvidia_p2p_override": {
+      "effective": true,
+      "configured": true,
+      "params_path": "/proc/driver/nvidia/params",
+      "params_available": true,
+      "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+      "modprobe_available": true,
+      "runtime": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1",
+        "DmaRemapPeerMmio": "1"
+      },
+      "expected": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1"
+      },
+      "missing": [],
+      "mismatched": {},
+      "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+      "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+      "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+    },
+    "p2pmark": {
+      "status": "not_run"
+    },
+    "amd_fabric": {
+      "status": "not_run"
+    },
+    "nvidia_smi_query": {
+      "cmd": [
+        "nvidia-smi",
+        "--query-gpu=index,name,driver_version,pci.bus_id,pcie.link.gen.current,pcie.link.width.current,power.limit",
+        "--format=csv,noheader,nounits"
+      ],
+      "returncode": 0,
+      "stdout": "0, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:01:00.0, 1, 16, 300.00\n1, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:21:00.0, 1, 16, 300.00\n2, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:81:00.0, 1, 16, 300.00\n3, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:C1:00.0, 1, 16, 300.00",
+      "stderr": ""
+    },
+    "nvidia_smi_topo": {
+      "cmd": [
+        "nvidia-smi",
+        "topo",
+        "-m"
+      ],
+      "returncode": 0,
+      "stdout": "\u001b[4mGPU0\tGPU1\tGPU2\tGPU3\tCPU Affinity\tNUMA Affinity\tGPU NUMA ID\u001b[0m\nGPU0\t X \tNODE\tNODE\tNODE\t0-47\t0\t\tN/A\nGPU1\tNODE\t X \tNODE\tNODE\t0-47\t0\t\tN/A\nGPU2\tNODE\tNODE\t X \tNODE\t0-47\t0\t\tN/A\nGPU3\tNODE\tNODE\tNODE\t X \t0-47\t0\t\tN/A\n\nLegend:\n\n  X    = Self\n  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)\n  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node\n  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)\n  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)\n  PIX  = Connection traversing at most a single PCIe bridge\n  NV#  = Connection traversing a bonded set of # NVLinks",
+      "stderr": ""
+    }
+  },
+  "nvidia_p2p_override": {
+    "effective": true,
+    "configured": true,
+    "params_path": "/proc/driver/nvidia/params",
+    "params_available": true,
+    "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+    "modprobe_available": true,
+    "runtime": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1",
+      "DmaRemapPeerMmio": "1"
+    },
+    "expected": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1"
+    },
+    "missing": [],
+    "mismatched": {},
+    "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+    "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+    "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+  },
+  "p2pmark": {
+    "status": "not_run"
+  },
+  "amd_fabric": {
+    "status": "not_run"
+  },
+  "hardware_run_summary": {
+    "samples": 20,
+    "duration_seconds": 45.551,
+    "gpu_count": 4,
+    "cpu_util_avg_pct": 9.75,
+    "cpu_temp_max_c": 76.5,
+    "gpu_util_avg_pct": 84.84,
+    "gpu_util_max_pct": 100.0,
+    "mem_util_avg_pct": 40.79,
+    "mem_util_max_pct": 62.0,
+    "temp_avg_c": 52.48,
+    "temp_max_c": 72.0,
+    "power_total_avg_w": 1017.83,
+    "power_total_max_w": 1173.63,
+    "power_limit_total_w": 1200.0,
+    "vram_used_avg_mb": 384370.0,
+    "vram_used_max_mb": 384370.0,
+    "vram_total_mb": 391548.0,
+    "vram_used_avg_pct": 98.17,
+    "vram_used_max_pct": 98.17,
+    "pcie_rx_avg_mb_s": 15291.5,
+    "pcie_rx_max_mb_s": 52252.0,
+    "pcie_tx_avg_mb_s": 14052.1,
+    "pcie_tx_max_mb_s": 44656.0
+  },
+  "event_log": [
+    "14:35:49 benchmark start engine=vllm",
+    "14:35:49 startup server=http://127.0.0.1:8001 model=glm53-flash-trellismx-p8-k45",
+    "14:35:49 startup decode concurrency=8 contexts=8k",
+    "14:35:49 startup NVIDIA P2P override: enabled: runtime NVIDIA P2P override matches expected RegistryDwords",
+    "14:35:49 startup engine vLLM 0.26.1rc0+glm53.flash.nvfp4.luke.clean.r1.vllme75bcfd.b12x58a046f models=['glm53-flash-trellismx-p8-k45']",
+    "14:35:49 startup KV cache budget from vLLM metrics: 29,532,160 tokens (3605 blocks x 2048; local 7,383,040 \u00d7 CP 4; CP source: local process)",
+    "14:35:49 startup model context length: 1,000,000 tokens",
+    "14:35:49 startup prefill tests: skipped",
+    "14:35:49 startup calibrating padding text run=tmxrepeatabc up_to=8k",
+    "14:35:49 startup context 8k: 50,558 chars (8,192 prompt tokens via /tokenize)",
+    "14:35:49 startup token targeting: /tokenize exact",
+    "14:35:49 startup startup preparation done",
+    "14:35:49 hardware monitor interval=2s",
+    "14:35:49 decode warmup start",
+    "14:35:49 decode warmup start C=1 ctx=8k 3s",
+    "14:35:49 cell start C=1 ctx=8k",
+    "14:35:54 ready C=1 ctx=8k running_reqs=1/1, queue_reqs=0, active_streams=1/1, stable=3.0s",
+    "14:35:57 cell done C=1 ctx=8k 224.3 tok/s",
+    "14:35:57 decode warmup done C=1 ctx=8k",
+    "14:35:59 cell start C=8 ctx=8k",
+    "14:36:12 ready C=8 ctx=8k running_reqs=8/8, queue_reqs=0, active_streams=8/8, stable=3.0s",
+    "14:36:33 cell done C=8 ctx=8k 594.6 tok/s"
+  ],
+  "prefill": {},
+  "results": [
+    {
+      "concurrency": 8,
+      "context_tokens": 8192,
+      "benchmark_mode": "duration",
+      "request_count_target": 0,
+      "warmup_request_count": 0,
+      "measurement_seconds": 19.995324,
+      "measurement_wall_seconds": 20.000426,
+      "client_output_tokens": 11890,
+      "server_output_tokens": 11890,
+      "aggregate_source": "openai_continuous_usage",
+      "aggregate_tps": 594.6390215769761,
+      "per_request_avg_tps": 74.32987769712201,
+      "ttft_avg": 4.443870959017659,
+      "ttft_p50": 4.351237794035114,
+      "ttft_p90": 7.407004370144568,
+      "ttft_p99": 8.707161026264075,
+      "time_to_second_token_avg": 0.03910974125028588,
+      "time_to_second_token_p50": 0.042902764049358666,
+      "time_to_second_token_p90": 0.04509180432651192,
+      "time_to_second_token_p99": 0.045149405833799394,
+      "request_latency_avg": 0.0,
+      "request_latency_p50": 0.0,
+      "request_latency_p90": 0.0,
+      "request_latency_p99": 0.0,
+      "inter_token_latency_avg": 0.014111235546572226,
+      "inter_token_latency_p50": 0.014077106559602584,
+      "inter_token_latency_p90": 0.014635102649253221,
+      "inter_token_latency_p99": 0.014888121111171413,
+      "output_tps_per_user_avg": 70.94257346566246,
+      "output_tps_per_user_p50": 71.03927256447412,
+      "output_tps_per_user_p90": 72.97073022440078,
+      "output_tps_per_user_p99": 75.48528683994984,
+      "e2e_output_tps_per_user_avg": 0.0,
+      "e2e_output_tps_per_user_p50": 0.0,
+      "e2e_output_tps_per_user_p90": 0.0,
+      "e2e_output_tps_per_user_p99": 0.0,
+      "chunk_inter_token_latency_avg": 0.03844512127484256,
+      "chunk_inter_token_latency_p50": 0.0388058107017188,
+      "chunk_inter_token_latency_p90": 0.0398085934959969,
+      "chunk_inter_token_latency_p99": 0.04017289479777741,
+      "input_seq_len_avg": 8192.0,
+      "output_seq_len_avg": 1993.875,
+      "output_seq_len_p50": 1971.0,
+      "output_seq_len_p90": 2228.2,
+      "output_seq_len_p99": 2287.42,
+      "request_count": 8,
+      "completed_request_count": 0,
+      "request_samples": [
+        {
+          "ttft": 8.85162287694402,
+          "time_to_second_token": 0.04145898902788758,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.013198761921065203,
+          "chunk_inter_token_latency_avg": 0.03628650448635357,
+          "input_tokens": 8192,
+          "output_tokens": 1799,
+          "output_tps_per_user": 75.7646820194553,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        },
+        {
+          "ttft": 0.5325793139636517,
+          "time_to_second_token": 0.01553312293253839,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.013977347604883287,
+          "chunk_inter_token_latency_avg": 0.04021337272019746,
+          "input_tokens": 8192,
+          "output_tokens": 2294,
+          "output_tps_per_user": 71.54433217720282,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        },
+        {
+          "ttft": 1.9443869879469275,
+          "time_to_second_token": 0.038359275087714195,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.013932753393380798,
+          "chunk_inter_token_latency_avg": 0.03953306414457339,
+          "input_tokens": 8192,
+          "output_tokens": 2200,
+          "output_tps_per_user": 71.77332231223457,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        },
+        {
+          "ttft": 1.9443444930948317,
+          "time_to_second_token": 0.038377634016796947,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.014916234273606768,
+          "chunk_inter_token_latency_avg": 0.03963511668562523,
+          "input_tokens": 8192,
+          "output_tokens": 2055,
+          "output_tps_per_user": 67.04104948052672,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        },
+        {
+          "ttft": 4.351254750043154,
+          "time_to_second_token": 0.04506437503732741,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.01415080536240922,
+          "chunk_inter_token_latency_avg": 0.03893911268690537,
+          "input_tokens": 8192,
+          "output_tokens": 1996,
+          "output_tps_per_user": 70.66735598359942,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        },
+        {
+          "ttft": 4.351220838027075,
+          "time_to_second_token": 0.04515580600127578,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.014514617667387415,
+          "chunk_inter_token_latency_avg": 0.038672508716532224,
+          "input_tokens": 8192,
+          "output_tokens": 1946,
+          "output_tps_per_user": 68.89606208828212,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        },
+        {
+          "ttft": 6.787882152944803,
+          "time_to_second_token": 0.04434653907082975,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.014195956393049176,
+          "chunk_inter_token_latency_avg": 0.03716722300600915,
+          "input_tokens": 8192,
+          "output_tokens": 1818,
+          "output_tps_per_user": 70.44259451864998,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        },
+        {
+          "ttft": 6.787676259176806,
+          "time_to_second_token": 0.04458218882791698,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.014003407756795948,
+          "chunk_inter_token_latency_avg": 0.037114067752544076,
+          "input_tokens": 8192,
+          "output_tokens": 1843,
+          "output_tps_per_user": 71.4111891453488,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        }
+      ],
+      "total_tokens": 11890,
+      "wall_time": 33.637119472026825,
+      "num_completed": 8,
+      "num_errors": 0,
+      "server_gen_throughput": 594.360504181179,
+      "server_utilization": 0.04883462819089901,
+      "server_spec_accept_rate": 0.5714285714285714,
+      "server_spec_accept_length": 0.0,
+      "avg_running_reqs": 8,
+      "max_running_reqs": 8,
+      "effective_concurrency": 8,
+      "avg_queue_reqs": 0,
+      "max_queue_reqs": 0,
+      "queue_fraction": 0.0,
+      "underfilled": false,
+      "warmup_timed_out": false,
+      "warmup_duration": 13.603,
+      "ready_reason": "running_reqs=8/8, queue_reqs=0, active_streams=8/8, stable=3.0s",
+      "timeout_reason": "",
+      "capacity_limited": false,
+      "hardware_summary": {
+        "samples": 9,
+        "duration_seconds": 19.215,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.82,
+        "cpu_temp_max_c": 76.5,
+        "gpu_util_avg_pct": 100.0,
+        "gpu_util_max_pct": 100.0,
+        "mem_util_avg_pct": 51.17,
+        "mem_util_max_pct": 59.0,
+        "temp_avg_c": 56.5,
+        "temp_max_c": 72.0,
+        "power_total_avg_w": 1169.91,
+        "power_total_max_w": 1173.63,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 14848.56,
+        "pcie_rx_max_mb_s": 16432.0,
+        "pcie_tx_avg_mb_s": 14370.89,
+        "pcie_tx_max_mb_s": 15808.0
+      }
+    }
+  ],
+  "summary_table": {
+    "8192": {
+      "8": 594.6390215769761
+    }
+  },
+  "burst_results": [],
+  "burst_summary_table": {},
+  "methodology": {
+    "prefill": {
+      "name": "Prefill",
+      "present": false,
+      "mode": "skipped",
+      "formula": "prompt_tokens / TTFT",
+      "notes": "Default mode records the required decode scout request for each non-zero decode context, so normal runs do not pay for a separate prefill phase. Standalone mode repeats cold-prefill samples. Prometheus prefill counters, when available and uncontaminated, are stored as validation."
+    },
+    "sustained_decode": {
+      "name": "Sustained Decode",
+      "present": true,
+      "formula": "OpenAI stream usage completion_tokens per measured window; client chunk fallback only when continuous usage is unavailable",
+      "notes": "Duration-based steady-state cell after warmup. This is the main tuning/regression signal for kernels, NCCL, DCP, MTP, and scheduling. Prometheus metrics are stored as validation and scheduler state, not the default headline."
+    },
+    "burst_e2e_decode": {
+      "name": "Burst / E2E Decode",
+      "present": false,
+      "status": "not run; use --run-burst",
+      "formula": "sum(completion_tokens) / profiling_wall_time",
+      "notes": "Finite client-facing request burst using OpenAI stream usage. It includes request admission, scheduling, prefill/cache behavior, and completion."
+    }
+  }
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/decode8k-c8.log

@@ -0,0 +1,109 @@
+{"diagnostic_wrapper": "fixed-prefix-greedy", "benchmark_sha256": "7239958032ec10781d7db3efca23a7602bd9aeb94455a723e2634ab7d6fe546a", "run_id": "tmxrepeatabc", "temperature": 0.0, "scope": "Separate diagnostic, not model-default throughput"}
+
+New version available: v0.6.2 (current: v0.4.29)
+Upgrade and restart? [Y/n]: Skipping update.
+
+╭──────────────────────────── NVIDIA P2P Override ─────────────────────────────╮
+│ Effective: yes                                                               │
+│ Configured file: yes (/etc/modprobe.d/nvidia-p2p-override.conf)              │
+│ Runtime: ForceP2P=0x11; RMForceP2PType=1; RMPcieP2PType=2;                   │
+│ GrdmaPciTopoCheckOverride=1; EnableResizableBar=1; DmaRemapPeerMmio=1        │
+╰──────────────────────────────────────────────────────────────────────────────╯
+╭─────────────────────────────── Configuration ────────────────────────────────╮
+│ LLM Inference Benchmark                                                      │
+│ Model: glm53-flash-trellismx-p8-k45 @ 127.0.0.1:8001                         │
+│ Decode concurrency: [8]                                                      │
+│ Decode contexts: ['8k']                                                      │
+│ Duration: 20.0s per decode test | Max tokens: 8192                           │
+│ Pre-decode warmup: C=1 max-runnable context for 3s                           │
+│ Prefill: skipped | Sustained decode: 1 cells                                 │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Engine: vLLM 0.26.1rc0+glm53.flash.nvfp4.luke.clean.r1.vllme75bcfd.b12x58a046f  
+Models: ['glm53-flash-trellismx-p8-k45']
+KV cache budget (vLLM metrics): 29,532,160 tokens (3605 blocks × 2048; local 
+7,383,040 × CP 4; CP source: local process)
+Model context length: 1,000,000 tokens
+Prefill tests: skipped
+Calibrating padding text (run=tmxrepeatabc, up to 8k)...
+  8k: 50,558 chars (8,192 prompt tokens via /tokenize)
+  Token targeting: /tokenize exact
+Done.
+
+
+
+llm-decode-bench v0.4.29
+╭────────────────────────────────── Phase 2 ───────────────────────────────────╮
+│ Sustained Decode                                                             │
+│ Steady-state decode throughput after the engine has admitted the requested   │
+│ concurrency and passed warmup. Use this as the main tuning/regression signal │
+│ for kernels, NCCL, DCP, MTP, and scheduler changes.                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Aggregate tok/s  + TTFT/ITL 
+╭────────────┬─────────────╮
+│ ctx \ conc │           8 │
+├────────────┼─────────────┤
+│ 8k         │ 594.6 4k/14 │
+╰────────────┴─────────────╯
+Sustained Decode: aggregate tok/s uses OpenAI stream usage by default 
+(continuous completion_tokens when the server supports it). Prometheus is kept 
+as validation/scheduler data.
+Aggregate source(s): openai_continuous_usage
+Per-Request tok/s    
+╭────────────┬──────╮
+│ ctx \ conc │    8 │
+├────────────┼──────┤
+│ 8k         │ 74.3 │
+╰────────────┴──────╯
+Client request      
+latency: p50 / p90  
+ms                  
+╭────────────┬─────╮
+│ ctx \ conc │   8 │
+├────────────┼─────┤
+│ 8k         │ —/— │
+╰────────────┴─────╯
+Aggregate cells show dim detail as TTFT ms / ITL ms for the same ctx/conc 
+coordinate. ITL is computed from observed generated tokens, including streams 
+stopped at the measurement boundary; a missing ITL means no stream produced at 
+least two measured output tokens. Per-request tok/s and request latency are 
+shown in separate per-cell matrices. Completion/sample counts and full 
+request-level distributions remain in JSON under request_samples.
+Sustained mode: client latency metrics explain request UX variance; aggregate 
+tok/s remains the primary throughput signal. 
+ITL=(last_token_time-first_token_time)/(output_tokens-1), user tok/s=1/ITL.
+Hardware Summary                                                                
+╭───┬─┬───────┬───────────┬───────┬─────────┬─────┬──────┬─────┬───────────────╮
+│ … │ │ mode  │ GPU avg/… │ Mem … │ W avg/… │ T … │ CPU… │ VR… │ PCIe rx/tx a… │
+├───┼─┼───────┼───────────┼───────┼─────────┼─────┼──────┼─────┼───────────────┤
+│ … │ │ sust… │  100/100% │   51% │ 1170/1… │ 72C │  76C │ 98… │   14849/14371 │
+╰───┴─┴───────┴───────────┴───────┴─────────┴─────┴──────┴─────┴───────────────╯
+╭─────────────────────── Whole-run GPU Power ───────────────────────╮
+│ avg 1,018 W | max 1,174 W | limit 1,200 W | over 46s | 20 samples │
+╰───────────────────────────────────────────────────────────────────╯
+Hardware summary is sampled from nvidia-smi during the measured part of each 
+cell. Whole-run GPU power is the sampled sum of GPU power draw across the 
+complete benchmark run, not wall-outlet system power. PCIe rx/tx is MB/s and is 
+a coarse live diagnostic, not a per-kernel NCCL profiler.
+
+╭────────────────────────────────── Phase 3 ───────────────────────────────────╮
+│ Burst / E2E Decode                                                           │
+│ Not run. Re-run with --run-burst to append a finite client-facing request    │
+│ burst after Sustained Decode. This is intentionally disabled by default      │
+│ because it adds another full decode matrix.                                  │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭────────────────────────────── Primary Summary ───────────────────────────────╮
+│ Primary matrices repeated last so the important numbers are visible without  │
+│ scrolling back through diagnostics.                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Aggregate decode tok/s
+╭────────────┬───────╮
+│ ctx \ conc │     8 │
+├────────────┼───────┤
+│ 8k         │ 594.6 │
+╰────────────┴───────╯
+
+Results saved to 
+<campaign>/tile-policy-route-
+hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/decode8k-c8.json
+{"fixed_run_id_calls": 1, "benchmark_unchanged": true}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill128k-command.json

@@ -0,0 +1,33 @@
+[
+  "/usr/bin/python3",
+  "<campaign>/fixed-workload-candidate-comparison-01/bench_fixed.py",
+  "--host",
+  "127.0.0.1",
+  "--port",
+  "8001",
+  "--model",
+  "glm53-flash-trellismx-p8-k45",
+  "--duration",
+  "20",
+  "--max-tokens",
+  "8192",
+  "--token-targeting",
+  "exact",
+  "--display-mode",
+  "plain",
+  "--output",
+  "<campaign>/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill128k.json",
+  "--contexts",
+  "8k",
+  "--concurrency",
+  "1",
+  "--prefill-only",
+  "--prefill-contexts",
+  "128k",
+  "--prefill-duration",
+  "20",
+  "--cell-warmup-timeout-seconds",
+  "180",
+  "--temperature",
+  "0"
+]

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill128k-cooldown.json

@@ -0,0 +1,84 @@
+{
+  "status": "passed",
+  "minimum_seconds": 90,
+  "temperature_ceiling_c": 55,
+  "stable_seconds": 30,
+  "samples": [
+    {
+      "elapsed_seconds": 2.2980384528636932e-07,
+      "hardware_csv": "0, 58, 2310, 115.24, 300.00\n1, 42, 2872, 134.79, 300.00\n2, 49, 2310, 109.57, 300.00\n3, 43, 2610, 119.54, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 5.104221327928826,
+      "hardware_csv": "0, 56, 2272, 75.58, 300.00\n1, 40, 2610, 86.86, 300.00\n2, 47, 2272, 73.62, 300.00\n3, 42, 2610, 97.96, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 10.21492264396511,
+      "hardware_csv": "0, 55, 2272, 74.07, 300.00\n1, 40, 2610, 86.21, 300.00\n2, 46, 2272, 72.89, 300.00\n3, 42, 2610, 97.00, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 15.316480945795774,
+      "hardware_csv": "0, 55, 2272, 74.27, 300.00\n1, 40, 2610, 86.35, 300.00\n2, 46, 2272, 72.26, 300.00\n3, 42, 2610, 97.07, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 20.427333840867504,
+      "hardware_csv": "0, 52, 180, 13.35, 300.00\n1, 38, 180, 28.78, 300.00\n2, 44, 180, 11.60, 300.00\n3, 40, 180, 25.33, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 25.57664366788231,
+      "hardware_csv": "0, 51, 172, 12.73, 300.00\n1, 37, 180, 15.53, 300.00\n2, 43, 180, 9.19, 300.00\n3, 39, 180, 20.33, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 30.707700602943078,
+      "hardware_csv": "0, 50, 180, 13.09, 300.00\n1, 37, 180, 15.75, 300.00\n2, 42, 180, 8.68, 300.00\n3, 39, 180, 22.72, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 35.84929059678689,
+      "hardware_csv": "0, 49, 180, 12.64, 300.00\n1, 36, 180, 15.53, 300.00\n2, 41, 180, 8.37, 300.00\n3, 38, 180, 19.66, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 40.98280342691578,
+      "hardware_csv": "0, 48, 180, 12.23, 300.00\n1, 36, 180, 15.82, 300.00\n2, 40, 180, 8.03, 300.00\n3, 38, 172, 21.65, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 46.115717354929075,
+      "hardware_csv": "0, 48, 180, 12.02, 300.00\n1, 35, 180, 15.44, 300.00\n2, 40, 180, 7.66, 300.00\n3, 37, 180, 19.66, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 51.2488560769707,
+      "hardware_csv": "0, 47, 180, 11.65, 300.00\n1, 35, 180, 15.35, 300.00\n2, 39, 180, 7.62, 300.00\n3, 37, 180, 19.03, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 56.384223312838,
+      "hardware_csv": "0, 47, 180, 11.55, 300.00\n1, 35, 180, 15.54, 300.00\n2, 38, 180, 7.60, 300.00\n3, 36, 180, 19.16, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 61.51466837897897,
+      "hardware_csv": "0, 46, 180, 10.89, 300.00\n1, 34, 172, 15.42, 300.00\n2, 38, 180, 7.61, 300.00\n3, 36, 180, 18.56, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 66.64880179893225,
+      "hardware_csv": "0, 46, 180, 10.69, 300.00\n1, 34, 180, 15.37, 300.00\n2, 37, 180, 7.48, 300.00\n3, 36, 180, 19.32, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 71.77847859985195,
+      "hardware_csv": "0, 45, 180, 10.57, 300.00\n1, 34, 180, 15.41, 300.00\n2, 37, 180, 6.78, 300.00\n3, 35, 180, 18.76, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 76.9130208818242,
+      "hardware_csv": "0, 45, 180, 10.71, 300.00\n1, 33, 180, 15.44, 300.00\n2, 37, 180, 6.95, 300.00\n3, 35, 180, 18.83, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 82.04496841598302,
+      "hardware_csv": "0, 44, 180, 10.17, 300.00\n1, 33, 180, 15.47, 300.00\n2, 36, 180, 6.99, 300.00\n3, 35, 180, 18.36, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 87.18223228678107,
+      "hardware_csv": "0, 44, 180, 9.87, 300.00\n1, 33, 180, 15.36, 300.00\n2, 36, 180, 7.05, 300.00\n3, 35, 180, 18.71, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 92.31365287397057,
+      "hardware_csv": "0, 43, 180, 9.96, 300.00\n1, 33, 180, 15.40, 300.00\n2, 35, 180, 6.89, 300.00\n3, 34, 180, 18.67, 300.00\n"
+    }
+  ]
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill128k-fixed-workload.json

@@ -0,0 +1,5 @@
+{
+  "prefix": "tmxrepeatabc",
+  "temperature": 0,
+  "scope": "greedy fixed-prompt diagnostic"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill128k-receipt.json

@@ -0,0 +1,5 @@
+{
+  "exit_code": 0,
+  "result_exists": true,
+  "sha256": "94fff041167a8bee4e516d8b1bb02a332b14584bc1203cb84dc0330c7c1a9f90"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill128k.json

@@ -0,0 +1,256 @@
+{
+  "metadata": {
+    "version": "0.4.29",
+    "engine": "vllm",
+    "model": "glm53-flash-trellismx-p8-k45",
+    "server": "127.0.0.1:8001",
+    "timestamp": "2026-09-09T14:19:57.566035",
+    "decode_mode": "duration",
+    "primary_decode_layer": "sustained_decode",
+    "duration_per_test": 20.0,
+    "request_count": 0,
+    "warmup_request_count": 0,
+    "run_burst": false,
+    "prefill_mode": "standalone_cold",
+    "standalone_prefill": true,
+    "prefill_only": true,
+    "skip_prefill": false,
+    "burst_e2e_status": "not_run_use_--run-burst",
+    "burst_request_count": 0,
+    "burst_warmup_request_count": 0,
+    "burst_requests_per_concurrency": 5,
+    "decode_warmup_seconds": 3.0,
+    "decode_warmup_context": 0,
+    "decode_warmup_concurrency": 1,
+    "cell_warmup_timeout_seconds": 180.0,
+    "cell_warmup_timeout_policy": "<=32k:60s,64k:120s,>=128k:180s when override is 0",
+    "show_capacity_limited_values": false,
+    "max_tokens": 8192,
+    "temperature": 0.0,
+    "ignore_eos": true,
+    "max_total_tokens": 29532160,
+    "dcp_size": 0,
+    "metrics_available": true,
+    "metrics_warning": "",
+    "concurrency_levels": [
+      1
+    ],
+    "context_lengths": [
+      8192
+    ],
+    "startup_diagnostics_available": true,
+    "nvidia_p2p_override_effective": true,
+    "p2pmark_status": "not_run",
+    "amd_fabric_status": "not_run"
+  },
+  "startup_diagnostics": {
+    "version": "0.4.29",
+    "server_url": "http://127.0.0.1:8001",
+    "hostname": "<host>",
+    "uname": "Linux <host> 6.18.7-76061807-generic #202601231045~1769703228~24.04~cb87b5b SMP PREEMPT_DYNAMIC Thu J x86_64 x86_64 x86_64 GNU/Linux",
+    "env": {},
+    "args": {
+      "concurrency": "1",
+      "contexts": "8k",
+      "max_tokens": 8192,
+      "duration": 20.0,
+      "request_count": 0,
+      "run_burst": false,
+      "standalone_prefill": true,
+      "prefill_only": true,
+      "skip_prefill": false,
+      "prefill_contexts": "128k",
+      "prefill_metric": "client",
+      "dcp_size": 0,
+      "kv_budget": 0
+    },
+    "nvidia_p2p_override": {
+      "effective": true,
+      "configured": true,
+      "params_path": "/proc/driver/nvidia/params",
+      "params_available": true,
+      "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+      "modprobe_available": true,
+      "runtime": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1",
+        "DmaRemapPeerMmio": "1"
+      },
+      "expected": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1"
+      },
+      "missing": [],
+      "mismatched": {},
+      "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+      "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+      "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+    },
+    "p2pmark": {
+      "status": "not_run"
+    },
+    "amd_fabric": {
+      "status": "not_run"
+    },
+    "nvidia_smi_query": {
+      "cmd": [
+        "nvidia-smi",
+        "--query-gpu=index,name,driver_version,pci.bus_id,pcie.link.gen.current,pcie.link.width.current,power.limit",
+        "--format=csv,noheader,nounits"
+      ],
+      "returncode": 0,
+      "stdout": "0, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:01:00.0, 1, 16, 300.00\n1, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:21:00.0, 1, 16, 300.00\n2, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:81:00.0, 1, 16, 300.00\n3, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:C1:00.0, 1, 16, 300.00",
+      "stderr": ""
+    },
+    "nvidia_smi_topo": {
+      "cmd": [
+        "nvidia-smi",
+        "topo",
+        "-m"
+      ],
+      "returncode": 0,
+      "stdout": "\u001b[4mGPU0\tGPU1\tGPU2\tGPU3\tCPU Affinity\tNUMA Affinity\tGPU NUMA ID\u001b[0m\nGPU0\t X \tNODE\tNODE\tNODE\t0-47\t0\t\tN/A\nGPU1\tNODE\t X \tNODE\tNODE\t0-47\t0\t\tN/A\nGPU2\tNODE\tNODE\t X \tNODE\t0-47\t0\t\tN/A\nGPU3\tNODE\tNODE\tNODE\t X \t0-47\t0\t\tN/A\n\nLegend:\n\n  X    = Self\n  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)\n  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node\n  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)\n  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)\n  PIX  = Connection traversing at most a single PCIe bridge\n  NV#  = Connection traversing a bonded set of # NVLinks",
+      "stderr": ""
+    }
+  },
+  "nvidia_p2p_override": {
+    "effective": true,
+    "configured": true,
+    "params_path": "/proc/driver/nvidia/params",
+    "params_available": true,
+    "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+    "modprobe_available": true,
+    "runtime": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1",
+      "DmaRemapPeerMmio": "1"
+    },
+    "expected": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1"
+    },
+    "missing": [],
+    "mismatched": {},
+    "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+    "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+    "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+  },
+  "p2pmark": {
+    "status": "not_run"
+  },
+  "amd_fabric": {
+    "status": "not_run"
+  },
+  "hardware_run_summary": {
+    "samples": 18,
+    "duration_seconds": 40.838,
+    "gpu_count": 4,
+    "cpu_util_avg_pct": 12.96,
+    "cpu_temp_max_c": 75.38,
+    "gpu_util_avg_pct": 88.88,
+    "gpu_util_max_pct": 100.0,
+    "mem_util_avg_pct": 24.15,
+    "mem_util_max_pct": 32.0,
+    "temp_avg_c": 51.03,
+    "temp_max_c": 68.0,
+    "power_total_avg_w": 1045.11,
+    "power_total_max_w": 1128.52,
+    "power_limit_total_w": 1200.0,
+    "vram_used_avg_mb": 384370.0,
+    "vram_used_max_mb": 384370.0,
+    "vram_total_mb": 391548.0,
+    "vram_used_avg_pct": 98.17,
+    "vram_used_max_pct": 98.17,
+    "pcie_rx_avg_mb_s": 48187.56,
+    "pcie_rx_max_mb_s": 60487.0,
+    "pcie_tx_avg_mb_s": 46382.72,
+    "pcie_tx_max_mb_s": 58392.0
+  },
+  "event_log": [],
+  "prefill": {
+    "131072": {
+      "ttft_seconds": 15.749,
+      "prefill_seconds": 15.749,
+      "tok_per_sec": 8323.0,
+      "client_ttft_seconds": 15.749,
+      "client_tok_per_sec": 8323.0,
+      "prompt_tokens": 131074,
+      "samples": 2,
+      "method": "client",
+      "server_validation": {
+        "method": "",
+        "tok_per_sec": 0.0,
+        "prefill_seconds": 0.0,
+        "prompt_tokens": 0,
+        "request_prompt_tokens": 0,
+        "cached_tokens": 0,
+        "token_source": "",
+        "samples": 0,
+        "invalid_reason": ""
+      },
+      "hardware_summary": {
+        "samples": 13,
+        "duration_seconds": 28.84,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.81,
+        "cpu_temp_max_c": 75.38,
+        "gpu_util_avg_pct": 99.98,
+        "gpu_util_max_pct": 100.0,
+        "mem_util_avg_pct": 27.08,
+        "mem_util_max_pct": 32.0,
+        "temp_avg_c": 53.13,
+        "temp_max_c": 68.0,
+        "power_total_avg_w": 1120.21,
+        "power_total_max_w": 1128.52,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 53419.31,
+        "pcie_rx_max_mb_s": 60487.0,
+        "pcie_tx_avg_mb_s": 51199.31,
+        "pcie_tx_max_mb_s": 58392.0
+      }
+    }
+  },
+  "results": [],
+  "summary_table": {},
+  "burst_results": [],
+  "burst_summary_table": {},
+  "methodology": {
+    "prefill": {
+      "name": "Prefill",
+      "present": true,
+      "mode": "standalone_cold",
+      "formula": "prompt_tokens / TTFT",
+      "notes": "Default mode records the required decode scout request for each non-zero decode context, so normal runs do not pay for a separate prefill phase. Standalone mode repeats cold-prefill samples. Prometheus prefill counters, when available and uncontaminated, are stored as validation."
+    },
+    "sustained_decode": {
+      "name": "Sustained Decode",
+      "present": false,
+      "formula": "OpenAI stream usage completion_tokens per measured window; client chunk fallback only when continuous usage is unavailable",
+      "notes": "Duration-based steady-state cell after warmup. This is the main tuning/regression signal for kernels, NCCL, DCP, MTP, and scheduling. Prometheus metrics are stored as validation and scheduler state, not the default headline."
+    },
+    "burst_e2e_decode": {
+      "name": "Burst / E2E Decode",
+      "present": false,
+      "status": "not run; use --run-burst",
+      "formula": "sum(completion_tokens) / profiling_wall_time",
+      "notes": "Finite client-facing request burst using OpenAI stream usage. It includes request admission, scheduling, prefill/cache behavior, and completion."
+    }
+  }
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill128k.log

@@ -0,0 +1,53 @@
+{"diagnostic_wrapper": "fixed-prefix-greedy", "benchmark_sha256": "7239958032ec10781d7db3efca23a7602bd9aeb94455a723e2634ab7d6fe546a", "run_id": "tmxrepeatabc", "temperature": 0.0, "scope": "Separate diagnostic, not model-default throughput"}
+
+New version available: v0.6.2 (current: v0.4.29)
+Upgrade and restart? [Y/n]: Skipping update.
+
+╭──────────────────────────── NVIDIA P2P Override ─────────────────────────────╮
+│ Effective: yes                                                               │
+│ Configured file: yes (/etc/modprobe.d/nvidia-p2p-override.conf)              │
+│ Runtime: ForceP2P=0x11; RMForceP2PType=1; RMPcieP2PType=2;                   │
+│ GrdmaPciTopoCheckOverride=1; EnableResizableBar=1; DmaRemapPeerMmio=1        │
+╰──────────────────────────────────────────────────────────────────────────────╯
+╭─────────────────────────────── Configuration ────────────────────────────────╮
+│ LLM Inference Benchmark                                                      │
+│ Model: glm53-flash-trellismx-p8-k45 @ 127.0.0.1:8001                         │
+│ Decode concurrency: [1]                                                      │
+│ Decode contexts: ['8k']                                                      │
+│ Decode: skipped (--prefill-only) | Max tokens: 8192                          │
+│ Pre-decode warmup: C=1 max-runnable context for 3s                           │
+│ Prefill-only: standalone cold profile (client) | Sustained decode: 0 cells   │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Engine: vLLM 0.26.1rc0+glm53.flash.nvfp4.luke.clean.r1.vllme75bcfd.b12x58a046f  
+Models: ['glm53-flash-trellismx-p8-k45']
+KV cache budget (vLLM metrics): 29,532,160 tokens (3605 blocks × 2048; local 
+7,383,040 × CP 4; CP source: local process)
+Model context length: 1,000,000 tokens
+Prefill tests: standalone cold profile ['128k']
+Calibrating padding text (run=tmxrepeatabc, up to 128k)...
+  8k: 50,558 chars (8,192 prompt tokens via /tokenize)
+  128k: 823,408 chars (131,072 prompt tokens via /tokenize)
+  Token targeting: /tokenize exact
+Done.
+
+
+
+llm-decode-bench v0.4.29
+Prefill Speed (C=1, client ISL / TTFT)                                          
+                                                                                
+                                                                PCIe rx/tx      
+  Context    Tokens   TTFT (s)   Client tok/s   Server tok/s           avg   N  
+ ────────────────────────────────────────────────────────────────────────────── 
+  128k      131,074      15.75          8,323              —   53419/51199   2  
+                                                                                
+Client tok/s = prompt_tokens / TTFT. Integrated scout rows come from the 
+prefix-cache scout request that decode needs anyway. Server tok/s is optional 
+Prometheus validation when the engine exports prefill counters and the exact 
+counter delta is uncontaminated; for vLLM this uses newly computed KV tokens, 
+not request prompt tokens.
+
+
+Results saved to 
+<campaign>/tile-policy-route-
+hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill128k.json
+{"fixed_run_id_calls": 1, "benchmark_unchanged": true}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill32k-command.json

@@ -0,0 +1,33 @@
+[
+  "/usr/bin/python3",
+  "<campaign>/fixed-workload-candidate-comparison-01/bench_fixed.py",
+  "--host",
+  "127.0.0.1",
+  "--port",
+  "8001",
+  "--model",
+  "glm53-flash-trellismx-p8-k45",
+  "--duration",
+  "20",
+  "--max-tokens",
+  "8192",
+  "--token-targeting",
+  "exact",
+  "--display-mode",
+  "plain",
+  "--output",
+  "<campaign>/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill32k.json",
+  "--contexts",
+  "8k",
+  "--concurrency",
+  "1",
+  "--prefill-only",
+  "--prefill-contexts",
+  "32k",
+  "--prefill-duration",
+  "20",
+  "--cell-warmup-timeout-seconds",
+  "180",
+  "--temperature",
+  "0"
+]

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill32k-cooldown.json

@@ -0,0 +1,84 @@
+{
+  "status": "passed",
+  "minimum_seconds": 90,
+  "temperature_ceiling_c": 55,
+  "stable_seconds": 30,
+  "samples": [
+    {
+      "elapsed_seconds": 2.300366759300232e-07,
+      "hardware_csv": "0, 49, 2250, 104.24, 300.00\n1, 43, 2812, 126.26, 300.00\n2, 41, 2220, 102.22, 300.00\n3, 43, 2610, 115.21, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 5.098315336043015,
+      "hardware_csv": "0, 48, 2272, 69.77, 300.00\n1, 37, 2610, 85.40, 300.00\n2, 40, 2272, 68.71, 300.00\n3, 38, 2610, 95.10, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 10.198127119103447,
+      "hardware_csv": "0, 48, 2272, 69.54, 300.00\n1, 37, 2610, 85.63, 300.00\n2, 40, 2272, 69.33, 300.00\n3, 38, 2610, 94.90, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 15.29899545200169,
+      "hardware_csv": "0, 48, 2272, 69.67, 300.00\n1, 37, 2610, 85.30, 300.00\n2, 40, 2272, 69.07, 300.00\n3, 38, 2610, 94.24, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 20.3999125729315,
+      "hardware_csv": "0, 46, 180, 15.53, 300.00\n1, 35, 180, 32.52, 300.00\n2, 39, 180, 12.05, 300.00\n3, 37, 172, 61.40, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 25.531402257969603,
+      "hardware_csv": "0, 45, 180, 10.44, 300.00\n1, 35, 180, 15.48, 300.00\n2, 38, 180, 7.95, 300.00\n3, 36, 180, 18.29, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 30.6658823271282,
+      "hardware_csv": "0, 45, 180, 10.12, 300.00\n1, 35, 180, 15.33, 300.00\n2, 38, 180, 7.89, 300.00\n3, 36, 180, 17.82, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 35.79796473495662,
+      "hardware_csv": "0, 44, 180, 9.99, 300.00\n1, 34, 180, 15.43, 300.00\n2, 37, 180, 7.66, 300.00\n3, 35, 180, 17.55, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 40.93220895715058,
+      "hardware_csv": "0, 44, 180, 10.26, 300.00\n1, 34, 180, 15.20, 300.00\n2, 36, 180, 7.80, 300.00\n3, 35, 180, 17.27, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 46.064513575984165,
+      "hardware_csv": "0, 43, 180, 9.52, 300.00\n1, 33, 180, 15.07, 300.00\n2, 36, 180, 7.39, 300.00\n3, 35, 180, 17.40, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 51.19570593093522,
+      "hardware_csv": "0, 43, 180, 9.64, 300.00\n1, 33, 180, 15.28, 300.00\n2, 36, 180, 7.30, 300.00\n3, 34, 180, 17.59, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 56.329856063006446,
+      "hardware_csv": "0, 43, 180, 9.38, 300.00\n1, 33, 180, 15.10, 300.00\n2, 35, 180, 7.02, 300.00\n3, 34, 180, 17.34, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 61.46232441510074,
+      "hardware_csv": "0, 42, 180, 9.60, 300.00\n1, 33, 180, 15.27, 300.00\n2, 35, 180, 6.99, 300.00\n3, 34, 180, 17.43, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 66.5961151169613,
+      "hardware_csv": "0, 42, 180, 9.22, 300.00\n1, 32, 180, 15.08, 300.00\n2, 35, 180, 7.14, 300.00\n3, 34, 180, 17.38, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 71.7273281160742,
+      "hardware_csv": "0, 42, 180, 8.95, 300.00\n1, 32, 180, 15.01, 300.00\n2, 34, 180, 6.93, 300.00\n3, 33, 180, 17.50, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 76.8605677341111,
+      "hardware_csv": "0, 41, 180, 9.16, 300.00\n1, 32, 180, 15.00, 300.00\n2, 34, 180, 7.20, 300.00\n3, 33, 180, 17.58, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 81.99244347307831,
+      "hardware_csv": "0, 41, 180, 8.84, 300.00\n1, 32, 180, 14.91, 300.00\n2, 34, 180, 7.43, 300.00\n3, 33, 180, 17.21, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 87.12506642797962,
+      "hardware_csv": "0, 41, 180, 8.85, 300.00\n1, 31, 180, 14.94, 300.00\n2, 33, 180, 6.60, 300.00\n3, 33, 180, 17.64, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 92.26148488814943,
+      "hardware_csv": "0, 40, 172, 9.17, 300.00\n1, 31, 180, 14.84, 300.00\n2, 33, 180, 6.61, 300.00\n3, 33, 180, 17.29, 300.00\n"
+    }
+  ]
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill32k-fixed-workload.json

@@ -0,0 +1,5 @@
+{
+  "prefix": "tmxrepeatabc",
+  "temperature": 0,
+  "scope": "greedy fixed-prompt diagnostic"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill32k-receipt.json

@@ -0,0 +1,5 @@
+{
+  "exit_code": 0,
+  "result_exists": true,
+  "sha256": "3073b37d3fecb0a0a9bc8e3ad222613e3876dc217693aead3f87eda086f3c062"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill32k.json

@@ -0,0 +1,256 @@
+{
+  "metadata": {
+    "version": "0.4.29",
+    "engine": "vllm",
+    "model": "glm53-flash-trellismx-p8-k45",
+    "server": "127.0.0.1:8001",
+    "timestamp": "2026-09-09T14:15:34.538411",
+    "decode_mode": "duration",
+    "primary_decode_layer": "sustained_decode",
+    "duration_per_test": 20.0,
+    "request_count": 0,
+    "warmup_request_count": 0,
+    "run_burst": false,
+    "prefill_mode": "standalone_cold",
+    "standalone_prefill": true,
+    "prefill_only": true,
+    "skip_prefill": false,
+    "burst_e2e_status": "not_run_use_--run-burst",
+    "burst_request_count": 0,
+    "burst_warmup_request_count": 0,
+    "burst_requests_per_concurrency": 5,
+    "decode_warmup_seconds": 3.0,
+    "decode_warmup_context": 0,
+    "decode_warmup_concurrency": 1,
+    "cell_warmup_timeout_seconds": 180.0,
+    "cell_warmup_timeout_policy": "<=32k:60s,64k:120s,>=128k:180s when override is 0",
+    "show_capacity_limited_values": false,
+    "max_tokens": 8192,
+    "temperature": 0.0,
+    "ignore_eos": true,
+    "max_total_tokens": 29532160,
+    "dcp_size": 0,
+    "metrics_available": true,
+    "metrics_warning": "",
+    "concurrency_levels": [
+      1
+    ],
+    "context_lengths": [
+      8192
+    ],
+    "startup_diagnostics_available": true,
+    "nvidia_p2p_override_effective": true,
+    "p2pmark_status": "not_run",
+    "amd_fabric_status": "not_run"
+  },
+  "startup_diagnostics": {
+    "version": "0.4.29",
+    "server_url": "http://127.0.0.1:8001",
+    "hostname": "<host>",
+    "uname": "Linux <host> 6.18.7-76061807-generic #202601231045~1769703228~24.04~cb87b5b SMP PREEMPT_DYNAMIC Thu J x86_64 x86_64 x86_64 GNU/Linux",
+    "env": {},
+    "args": {
+      "concurrency": "1",
+      "contexts": "8k",
+      "max_tokens": 8192,
+      "duration": 20.0,
+      "request_count": 0,
+      "run_burst": false,
+      "standalone_prefill": true,
+      "prefill_only": true,
+      "skip_prefill": false,
+      "prefill_contexts": "32k",
+      "prefill_metric": "client",
+      "dcp_size": 0,
+      "kv_budget": 0
+    },
+    "nvidia_p2p_override": {
+      "effective": true,
+      "configured": true,
+      "params_path": "/proc/driver/nvidia/params",
+      "params_available": true,
+      "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+      "modprobe_available": true,
+      "runtime": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1",
+        "DmaRemapPeerMmio": "1"
+      },
+      "expected": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1"
+      },
+      "missing": [],
+      "mismatched": {},
+      "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+      "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+      "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+    },
+    "p2pmark": {
+      "status": "not_run"
+    },
+    "amd_fabric": {
+      "status": "not_run"
+    },
+    "nvidia_smi_query": {
+      "cmd": [
+        "nvidia-smi",
+        "--query-gpu=index,name,driver_version,pci.bus_id,pcie.link.gen.current,pcie.link.width.current,power.limit",
+        "--format=csv,noheader,nounits"
+      ],
+      "returncode": 0,
+      "stdout": "0, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:01:00.0, 1, 16, 300.00\n1, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:21:00.0, 1, 16, 300.00\n2, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:81:00.0, 1, 16, 300.00\n3, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:C1:00.0, 1, 16, 300.00",
+      "stderr": ""
+    },
+    "nvidia_smi_topo": {
+      "cmd": [
+        "nvidia-smi",
+        "topo",
+        "-m"
+      ],
+      "returncode": 0,
+      "stdout": "\u001b[4mGPU0\tGPU1\tGPU2\tGPU3\tCPU Affinity\tNUMA Affinity\tGPU NUMA ID\u001b[0m\nGPU0\t X \tNODE\tNODE\tNODE\t0-47\t0\t\tN/A\nGPU1\tNODE\t X \tNODE\tNODE\t0-47\t0\t\tN/A\nGPU2\tNODE\tNODE\t X \tNODE\t0-47\t0\t\tN/A\nGPU3\tNODE\tNODE\tNODE\t X \t0-47\t0\t\tN/A\n\nLegend:\n\n  X    = Self\n  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)\n  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node\n  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)\n  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)\n  PIX  = Connection traversing at most a single PCIe bridge\n  NV#  = Connection traversing a bonded set of # NVLinks",
+      "stderr": ""
+    }
+  },
+  "nvidia_p2p_override": {
+    "effective": true,
+    "configured": true,
+    "params_path": "/proc/driver/nvidia/params",
+    "params_available": true,
+    "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+    "modprobe_available": true,
+    "runtime": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1",
+      "DmaRemapPeerMmio": "1"
+    },
+    "expected": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1"
+    },
+    "missing": [],
+    "mismatched": {},
+    "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+    "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+    "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+  },
+  "p2pmark": {
+    "status": "not_run"
+  },
+  "amd_fabric": {
+    "status": "not_run"
+  },
+  "hardware_run_summary": {
+    "samples": 11,
+    "duration_seconds": 23.994,
+    "gpu_count": 4,
+    "cpu_util_avg_pct": 9.75,
+    "cpu_temp_max_c": 75.25,
+    "gpu_util_avg_pct": 86.89,
+    "gpu_util_max_pct": 100.0,
+    "mem_util_avg_pct": 24.86,
+    "mem_util_max_pct": 36.0,
+    "temp_avg_c": 46.61,
+    "temp_max_c": 62.0,
+    "power_total_avg_w": 1011.54,
+    "power_total_max_w": 1135.15,
+    "power_limit_total_w": 1200.0,
+    "vram_used_avg_mb": 383671.82,
+    "vram_used_max_mb": 384370.0,
+    "vram_total_mb": 391548.0,
+    "vram_used_avg_pct": 97.99,
+    "vram_used_max_pct": 98.17,
+    "pcie_rx_avg_mb_s": 46115.27,
+    "pcie_rx_max_mb_s": 65384.0,
+    "pcie_tx_avg_mb_s": 41433.64,
+    "pcie_tx_max_mb_s": 58878.0
+  },
+  "event_log": [],
+  "prefill": {
+    "32768": {
+      "ttft_seconds": 3.875,
+      "prefill_seconds": 3.875,
+      "tok_per_sec": 8457.0,
+      "client_ttft_seconds": 3.875,
+      "client_tok_per_sec": 8457.0,
+      "prompt_tokens": 32770,
+      "samples": 5,
+      "method": "client",
+      "server_validation": {
+        "method": "",
+        "tok_per_sec": 0.0,
+        "prefill_seconds": 0.0,
+        "prompt_tokens": 0,
+        "request_prompt_tokens": 0,
+        "cached_tokens": 0,
+        "token_source": "",
+        "samples": 0,
+        "invalid_reason": ""
+      },
+      "hardware_summary": {
+        "samples": 9,
+        "duration_seconds": 19.188,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.71,
+        "cpu_temp_max_c": 75.25,
+        "gpu_util_avg_pct": 95.08,
+        "gpu_util_max_pct": 100.0,
+        "mem_util_avg_pct": 27.33,
+        "mem_util_max_pct": 36.0,
+        "temp_avg_c": 48.33,
+        "temp_max_c": 62.0,
+        "power_total_avg_w": 1104.35,
+        "power_total_max_w": 1135.15,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 47818.11,
+        "pcie_rx_max_mb_s": 65384.0,
+        "pcie_tx_avg_mb_s": 41943.22,
+        "pcie_tx_max_mb_s": 54316.0
+      }
+    }
+  },
+  "results": [],
+  "summary_table": {},
+  "burst_results": [],
+  "burst_summary_table": {},
+  "methodology": {
+    "prefill": {
+      "name": "Prefill",
+      "present": true,
+      "mode": "standalone_cold",
+      "formula": "prompt_tokens / TTFT",
+      "notes": "Default mode records the required decode scout request for each non-zero decode context, so normal runs do not pay for a separate prefill phase. Standalone mode repeats cold-prefill samples. Prometheus prefill counters, when available and uncontaminated, are stored as validation."
+    },
+    "sustained_decode": {
+      "name": "Sustained Decode",
+      "present": false,
+      "formula": "OpenAI stream usage completion_tokens per measured window; client chunk fallback only when continuous usage is unavailable",
+      "notes": "Duration-based steady-state cell after warmup. This is the main tuning/regression signal for kernels, NCCL, DCP, MTP, and scheduling. Prometheus metrics are stored as validation and scheduler state, not the default headline."
+    },
+    "burst_e2e_decode": {
+      "name": "Burst / E2E Decode",
+      "present": false,
+      "status": "not run; use --run-burst",
+      "formula": "sum(completion_tokens) / profiling_wall_time",
+      "notes": "Finite client-facing request burst using OpenAI stream usage. It includes request admission, scheduling, prefill/cache behavior, and completion."
+    }
+  }
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill32k.log

@@ -0,0 +1,53 @@
+{"diagnostic_wrapper": "fixed-prefix-greedy", "benchmark_sha256": "7239958032ec10781d7db3efca23a7602bd9aeb94455a723e2634ab7d6fe546a", "run_id": "tmxrepeatabc", "temperature": 0.0, "scope": "Separate diagnostic, not model-default throughput"}
+
+New version available: v0.6.2 (current: v0.4.29)
+Upgrade and restart? [Y/n]: Skipping update.
+
+╭──────────────────────────── NVIDIA P2P Override ─────────────────────────────╮
+│ Effective: yes                                                               │
+│ Configured file: yes (/etc/modprobe.d/nvidia-p2p-override.conf)              │
+│ Runtime: ForceP2P=0x11; RMForceP2PType=1; RMPcieP2PType=2;                   │
+│ GrdmaPciTopoCheckOverride=1; EnableResizableBar=1; DmaRemapPeerMmio=1        │
+╰──────────────────────────────────────────────────────────────────────────────╯
+╭─────────────────────────────── Configuration ────────────────────────────────╮
+│ LLM Inference Benchmark                                                      │
+│ Model: glm53-flash-trellismx-p8-k45 @ 127.0.0.1:8001                         │
+│ Decode concurrency: [1]                                                      │
+│ Decode contexts: ['8k']                                                      │
+│ Decode: skipped (--prefill-only) | Max tokens: 8192                          │
+│ Pre-decode warmup: C=1 max-runnable context for 3s                           │
+│ Prefill-only: standalone cold profile (client) | Sustained decode: 0 cells   │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Engine: vLLM 0.26.1rc0+glm53.flash.nvfp4.luke.clean.r1.vllme75bcfd.b12x58a046f  
+Models: ['glm53-flash-trellismx-p8-k45']
+KV cache budget (vLLM metrics): 29,532,160 tokens (3605 blocks × 2048; local 
+7,383,040 × CP 4; CP source: local process)
+Model context length: 1,000,000 tokens
+Prefill tests: standalone cold profile ['32k']
+Calibrating padding text (run=tmxrepeatabc, up to 32k)...
+  8k: 50,558 chars (8,192 prompt tokens via /tokenize)
+  32k: 205,152 chars (32,768 prompt tokens via /tokenize)
+  Token targeting: /tokenize exact
+Done.
+
+
+
+llm-decode-bench v0.4.29
+Prefill Speed (C=1, client ISL / TTFT)                                          
+                                                                                
+                                                                PCIe rx/tx      
+  Context   Tokens   TTFT (s)   Client tok/s   Server tok/s            avg   N  
+ ────────────────────────────────────────────────────────────────────────────── 
+  32k       32,770       3.87          8,457              —    47818/41943   5  
+                                                                                
+Client tok/s = prompt_tokens / TTFT. Integrated scout rows come from the 
+prefix-cache scout request that decode needs anyway. Server tok/s is optional 
+Prometheus validation when the engine exports prefill counters and the exact 
+counter delta is uncontaminated; for vLLM this uses newly computed KV tokens, 
+not request prompt tokens.
+
+
+Results saved to 
+<campaign>/tile-policy-route-
+hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill32k.json
+{"fixed_run_id_calls": 1, "benchmark_unchanged": true}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill64k-command.json

@@ -0,0 +1,33 @@
+[
+  "/usr/bin/python3",
+  "<campaign>/fixed-workload-candidate-comparison-01/bench_fixed.py",
+  "--host",
+  "127.0.0.1",
+  "--port",
+  "8001",
+  "--model",
+  "glm53-flash-trellismx-p8-k45",
+  "--duration",
+  "20",
+  "--max-tokens",
+  "8192",
+  "--token-targeting",
+  "exact",
+  "--display-mode",
+  "plain",
+  "--output",
+  "<campaign>/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill64k.json",
+  "--contexts",
+  "8k",
+  "--concurrency",
+  "1",
+  "--prefill-only",
+  "--prefill-contexts",
+  "64k",
+  "--prefill-duration",
+  "20",
+  "--cell-warmup-timeout-seconds",
+  "180",
+  "--temperature",
+  "0"
+]

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill64k-cooldown.json

@@ -0,0 +1,84 @@
+{
+  "status": "passed",
+  "minimum_seconds": 90,
+  "temperature_ceiling_c": 55,
+  "stable_seconds": 30,
+  "samples": [
+    {
+      "elapsed_seconds": 2.9010698199272156e-07,
+      "hardware_csv": "0, 55, 2302, 102.49, 300.00\n1, 41, 2872, 123.95, 300.00\n2, 47, 2310, 82.05, 300.00\n3, 42, 2865, 154.51, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 5.107993258163333,
+      "hardware_csv": "0, 54, 2272, 73.39, 300.00\n1, 39, 2610, 85.62, 300.00\n2, 45, 2272, 71.71, 300.00\n3, 41, 2610, 96.92, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 10.204473510151729,
+      "hardware_csv": "0, 53, 2272, 72.46, 300.00\n1, 39, 2610, 85.56, 300.00\n2, 45, 2272, 71.28, 300.00\n3, 40, 2610, 96.32, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 15.31410523597151,
+      "hardware_csv": "0, 53, 2272, 72.04, 300.00\n1, 39, 2610, 85.48, 300.00\n2, 44, 2272, 71.01, 300.00\n3, 40, 2610, 96.13, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 20.42270627198741,
+      "hardware_csv": "0, 50, 180, 12.23, 300.00\n1, 37, 180, 17.08, 300.00\n2, 42, 180, 8.75, 300.00\n3, 38, 180, 32.98, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 25.56934497016482,
+      "hardware_csv": "0, 49, 180, 12.16, 300.00\n1, 36, 180, 15.37, 300.00\n2, 41, 180, 8.11, 300.00\n3, 38, 180, 22.14, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 30.70141045213677,
+      "hardware_csv": "0, 48, 180, 11.81, 300.00\n1, 36, 180, 15.10, 300.00\n2, 40, 180, 8.19, 300.00\n3, 37, 180, 19.82, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 35.836296811001375,
+      "hardware_csv": "0, 48, 180, 11.14, 300.00\n1, 35, 180, 16.02, 300.00\n2, 40, 180, 7.60, 300.00\n3, 37, 180, 19.40, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 40.96954852202907,
+      "hardware_csv": "0, 47, 180, 10.85, 300.00\n1, 35, 180, 15.32, 300.00\n2, 39, 180, 7.76, 300.00\n3, 37, 180, 19.08, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 46.11254705814645,
+      "hardware_csv": "0, 46, 180, 10.97, 300.00\n1, 35, 180, 16.02, 300.00\n2, 38, 180, 7.83, 300.00\n3, 36, 180, 19.35, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 51.24383017513901,
+      "hardware_csv": "0, 46, 180, 10.75, 300.00\n1, 34, 180, 15.15, 300.00\n2, 38, 180, 7.66, 300.00\n3, 36, 180, 18.83, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 56.379540583118796,
+      "hardware_csv": "0, 45, 172, 10.64, 300.00\n1, 34, 180, 15.22, 300.00\n2, 37, 180, 7.20, 300.00\n3, 35, 180, 18.65, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 61.51068438705988,
+      "hardware_csv": "0, 45, 180, 10.24, 300.00\n1, 34, 180, 15.85, 300.00\n2, 37, 180, 7.33, 300.00\n3, 35, 180, 18.37, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 66.65063258307055,
+      "hardware_csv": "0, 44, 180, 10.54, 300.00\n1, 33, 180, 15.19, 300.00\n2, 36, 180, 7.15, 300.00\n3, 35, 180, 18.78, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 71.78509690612555,
+      "hardware_csv": "0, 44, 180, 10.01, 300.00\n1, 33, 180, 15.07, 300.00\n2, 36, 180, 7.02, 300.00\n3, 35, 180, 18.41, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 76.9198347791098,
+      "hardware_csv": "0, 43, 180, 9.78, 300.00\n1, 33, 180, 15.08, 300.00\n2, 36, 180, 7.00, 300.00\n3, 34, 180, 18.72, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 82.06185967312194,
+      "hardware_csv": "0, 43, 180, 10.00, 300.00\n1, 32, 180, 15.35, 300.00\n2, 35, 180, 6.90, 300.00\n3, 34, 180, 17.92, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 87.20084064803086,
+      "hardware_csv": "0, 43, 180, 9.30, 300.00\n1, 32, 180, 15.14, 300.00\n2, 35, 180, 6.48, 300.00\n3, 34, 180, 18.74, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 92.32916426914744,
+      "hardware_csv": "0, 42, 180, 9.15, 300.00\n1, 32, 180, 15.18, 300.00\n2, 34, 180, 6.41, 300.00\n3, 34, 180, 18.21, 300.00\n"
+    }
+  ]
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill64k-fixed-workload.json

@@ -0,0 +1,5 @@
+{
+  "prefix": "tmxrepeatabc",
+  "temperature": 0,
+  "scope": "greedy fixed-prompt diagnostic"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill64k-receipt.json

@@ -0,0 +1,5 @@
+{
+  "exit_code": 0,
+  "result_exists": true,
+  "sha256": "50f4cf7547092b0622c52315150dc4ee1c4cb45268fbb6d5c3fe8b6b0515efbf"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill64k.json

@@ -0,0 +1,256 @@
+{
+  "metadata": {
+    "version": "0.4.29",
+    "engine": "vllm",
+    "model": "glm53-flash-trellismx-p8-k45",
+    "server": "127.0.0.1:8001",
+    "timestamp": "2026-09-09T14:17:39.160990",
+    "decode_mode": "duration",
+    "primary_decode_layer": "sustained_decode",
+    "duration_per_test": 20.0,
+    "request_count": 0,
+    "warmup_request_count": 0,
+    "run_burst": false,
+    "prefill_mode": "standalone_cold",
+    "standalone_prefill": true,
+    "prefill_only": true,
+    "skip_prefill": false,
+    "burst_e2e_status": "not_run_use_--run-burst",
+    "burst_request_count": 0,
+    "burst_warmup_request_count": 0,
+    "burst_requests_per_concurrency": 5,
+    "decode_warmup_seconds": 3.0,
+    "decode_warmup_context": 0,
+    "decode_warmup_concurrency": 1,
+    "cell_warmup_timeout_seconds": 180.0,
+    "cell_warmup_timeout_policy": "<=32k:60s,64k:120s,>=128k:180s when override is 0",
+    "show_capacity_limited_values": false,
+    "max_tokens": 8192,
+    "temperature": 0.0,
+    "ignore_eos": true,
+    "max_total_tokens": 29532160,
+    "dcp_size": 0,
+    "metrics_available": true,
+    "metrics_warning": "",
+    "concurrency_levels": [
+      1
+    ],
+    "context_lengths": [
+      8192
+    ],
+    "startup_diagnostics_available": true,
+    "nvidia_p2p_override_effective": true,
+    "p2pmark_status": "not_run",
+    "amd_fabric_status": "not_run"
+  },
+  "startup_diagnostics": {
+    "version": "0.4.29",
+    "server_url": "http://127.0.0.1:8001",
+    "hostname": "<host>",
+    "uname": "Linux <host> 6.18.7-76061807-generic #202601231045~1769703228~24.04~cb87b5b SMP PREEMPT_DYNAMIC Thu J x86_64 x86_64 x86_64 GNU/Linux",
+    "env": {},
+    "args": {
+      "concurrency": "1",
+      "contexts": "8k",
+      "max_tokens": 8192,
+      "duration": 20.0,
+      "request_count": 0,
+      "run_burst": false,
+      "standalone_prefill": true,
+      "prefill_only": true,
+      "skip_prefill": false,
+      "prefill_contexts": "64k",
+      "prefill_metric": "client",
+      "dcp_size": 0,
+      "kv_budget": 0
+    },
+    "nvidia_p2p_override": {
+      "effective": true,
+      "configured": true,
+      "params_path": "/proc/driver/nvidia/params",
+      "params_available": true,
+      "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+      "modprobe_available": true,
+      "runtime": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1",
+        "DmaRemapPeerMmio": "1"
+      },
+      "expected": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1"
+      },
+      "missing": [],
+      "mismatched": {},
+      "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+      "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+      "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+    },
+    "p2pmark": {
+      "status": "not_run"
+    },
+    "amd_fabric": {
+      "status": "not_run"
+    },
+    "nvidia_smi_query": {
+      "cmd": [
+        "nvidia-smi",
+        "--query-gpu=index,name,driver_version,pci.bus_id,pcie.link.gen.current,pcie.link.width.current,power.limit",
+        "--format=csv,noheader,nounits"
+      ],
+      "returncode": 0,
+      "stdout": "0, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:01:00.0, 1, 16, 300.00\n1, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:21:00.0, 1, 16, 300.00\n2, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:81:00.0, 1, 16, 300.00\n3, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:C1:00.0, 1, 16, 300.00",
+      "stderr": ""
+    },
+    "nvidia_smi_topo": {
+      "cmd": [
+        "nvidia-smi",
+        "topo",
+        "-m"
+      ],
+      "returncode": 0,
+      "stdout": "\u001b[4mGPU0\tGPU1\tGPU2\tGPU3\tCPU Affinity\tNUMA Affinity\tGPU NUMA ID\u001b[0m\nGPU0\t X \tNODE\tNODE\tNODE\t0-47\t0\t\tN/A\nGPU1\tNODE\t X \tNODE\tNODE\t0-47\t0\t\tN/A\nGPU2\tNODE\tNODE\t X \tNODE\t0-47\t0\t\tN/A\nGPU3\tNODE\tNODE\tNODE\t X \t0-47\t0\t\tN/A\n\nLegend:\n\n  X    = Self\n  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)\n  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node\n  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)\n  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)\n  PIX  = Connection traversing at most a single PCIe bridge\n  NV#  = Connection traversing a bonded set of # NVLinks",
+      "stderr": ""
+    }
+  },
+  "nvidia_p2p_override": {
+    "effective": true,
+    "configured": true,
+    "params_path": "/proc/driver/nvidia/params",
+    "params_available": true,
+    "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+    "modprobe_available": true,
+    "runtime": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1",
+      "DmaRemapPeerMmio": "1"
+    },
+    "expected": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1"
+    },
+    "missing": [],
+    "mismatched": {},
+    "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+    "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+    "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+  },
+  "p2pmark": {
+    "status": "not_run"
+  },
+  "amd_fabric": {
+    "status": "not_run"
+  },
+  "hardware_run_summary": {
+    "samples": 13,
+    "duration_seconds": 28.782,
+    "gpu_count": 4,
+    "cpu_util_avg_pct": 9.89,
+    "cpu_temp_max_c": 74.12,
+    "gpu_util_avg_pct": 86.15,
+    "gpu_util_max_pct": 100.0,
+    "mem_util_avg_pct": 24.12,
+    "mem_util_max_pct": 36.0,
+    "temp_avg_c": 48.56,
+    "temp_max_c": 65.0,
+    "power_total_avg_w": 1014.98,
+    "power_total_max_w": 1129.79,
+    "power_limit_total_w": 1200.0,
+    "vram_used_avg_mb": 384370.0,
+    "vram_used_max_mb": 384370.0,
+    "vram_total_mb": 391548.0,
+    "vram_used_avg_pct": 98.17,
+    "vram_used_max_pct": 98.17,
+    "pcie_rx_avg_mb_s": 44191.38,
+    "pcie_rx_max_mb_s": 60865.0,
+    "pcie_tx_avg_mb_s": 45035.69,
+    "pcie_tx_max_mb_s": 61757.0
+  },
+  "event_log": [],
+  "prefill": {
+    "65536": {
+      "ttft_seconds": 7.762,
+      "prefill_seconds": 7.762,
+      "tok_per_sec": 8443.0,
+      "client_ttft_seconds": 7.762,
+      "client_tok_per_sec": 8443.0,
+      "prompt_tokens": 65538,
+      "samples": 3,
+      "method": "client",
+      "server_validation": {
+        "method": "",
+        "tok_per_sec": 0.0,
+        "prefill_seconds": 0.0,
+        "prompt_tokens": 0,
+        "request_prompt_tokens": 0,
+        "cached_tokens": 0,
+        "token_source": "",
+        "samples": 0,
+        "invalid_reason": ""
+      },
+      "hardware_summary": {
+        "samples": 10,
+        "duration_seconds": 21.613,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.76,
+        "cpu_temp_max_c": 74.12,
+        "gpu_util_avg_pct": 96.5,
+        "gpu_util_max_pct": 100.0,
+        "mem_util_avg_pct": 26.88,
+        "mem_util_max_pct": 35.0,
+        "temp_avg_c": 49.95,
+        "temp_max_c": 65.0,
+        "power_total_avg_w": 1095.48,
+        "power_total_max_w": 1129.79,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 51563.0,
+        "pcie_rx_max_mb_s": 60865.0,
+        "pcie_tx_avg_mb_s": 51753.4,
+        "pcie_tx_max_mb_s": 61757.0
+      }
+    }
+  },
+  "results": [],
+  "summary_table": {},
+  "burst_results": [],
+  "burst_summary_table": {},
+  "methodology": {
+    "prefill": {
+      "name": "Prefill",
+      "present": true,
+      "mode": "standalone_cold",
+      "formula": "prompt_tokens / TTFT",
+      "notes": "Default mode records the required decode scout request for each non-zero decode context, so normal runs do not pay for a separate prefill phase. Standalone mode repeats cold-prefill samples. Prometheus prefill counters, when available and uncontaminated, are stored as validation."
+    },
+    "sustained_decode": {
+      "name": "Sustained Decode",
+      "present": false,
+      "formula": "OpenAI stream usage completion_tokens per measured window; client chunk fallback only when continuous usage is unavailable",
+      "notes": "Duration-based steady-state cell after warmup. This is the main tuning/regression signal for kernels, NCCL, DCP, MTP, and scheduling. Prometheus metrics are stored as validation and scheduler state, not the default headline."
+    },
+    "burst_e2e_decode": {
+      "name": "Burst / E2E Decode",
+      "present": false,
+      "status": "not run; use --run-burst",
+      "formula": "sum(completion_tokens) / profiling_wall_time",
+      "notes": "Finite client-facing request burst using OpenAI stream usage. It includes request admission, scheduling, prefill/cache behavior, and completion."
+    }
+  }
+}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill64k.log

@@ -0,0 +1,53 @@
+{"diagnostic_wrapper": "fixed-prefix-greedy", "benchmark_sha256": "7239958032ec10781d7db3efca23a7602bd9aeb94455a723e2634ab7d6fe546a", "run_id": "tmxrepeatabc", "temperature": 0.0, "scope": "Separate diagnostic, not model-default throughput"}
+
+New version available: v0.6.2 (current: v0.4.29)
+Upgrade and restart? [Y/n]: Skipping update.
+
+╭──────────────────────────── NVIDIA P2P Override ─────────────────────────────╮
+│ Effective: yes                                                               │
+│ Configured file: yes (/etc/modprobe.d/nvidia-p2p-override.conf)              │
+│ Runtime: ForceP2P=0x11; RMForceP2PType=1; RMPcieP2PType=2;                   │
+│ GrdmaPciTopoCheckOverride=1; EnableResizableBar=1; DmaRemapPeerMmio=1        │
+╰──────────────────────────────────────────────────────────────────────────────╯
+╭─────────────────────────────── Configuration ────────────────────────────────╮
+│ LLM Inference Benchmark                                                      │
+│ Model: glm53-flash-trellismx-p8-k45 @ 127.0.0.1:8001                         │
+│ Decode concurrency: [1]                                                      │
+│ Decode contexts: ['8k']                                                      │
+│ Decode: skipped (--prefill-only) | Max tokens: 8192                          │
+│ Pre-decode warmup: C=1 max-runnable context for 3s                           │
+│ Prefill-only: standalone cold profile (client) | Sustained decode: 0 cells   │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Engine: vLLM 0.26.1rc0+glm53.flash.nvfp4.luke.clean.r1.vllme75bcfd.b12x58a046f  
+Models: ['glm53-flash-trellismx-p8-k45']
+KV cache budget (vLLM metrics): 29,532,160 tokens (3605 blocks × 2048; local 
+7,383,040 × CP 4; CP source: local process)
+Model context length: 1,000,000 tokens
+Prefill tests: standalone cold profile ['64k']
+Calibrating padding text (run=tmxrepeatabc, up to 64k)...
+  8k: 50,558 chars (8,192 prompt tokens via /tokenize)
+  64k: 411,264 chars (65,536 prompt tokens via /tokenize)
+  Token targeting: /tokenize exact
+Done.
+
+
+
+llm-decode-bench v0.4.29
+Prefill Speed (C=1, client ISL / TTFT)                                          
+                                                                                
+                                                                PCIe rx/tx      
+  Context   Tokens   TTFT (s)   Client tok/s   Server tok/s            avg   N  
+ ────────────────────────────────────────────────────────────────────────────── 
+  64k       65,538       7.76          8,443              —    51563/51753   3  
+                                                                                
+Client tok/s = prompt_tokens / TTFT. Integrated scout rows come from the 
+prefix-cache scout request that decode needs anyway. Server tok/s is optional 
+Prometheus validation when the engine exports prefill counters and the exact 
+counter delta is uncontaminated; for vLLM this uses newly computed KV tokens, 
+not request prompt tokens.
+
+
+Results saved to 
+<campaign>/tile-policy-route-
+hoist-expanded-02/results-01/tile-policy-route-hoist/rep-1/prefill64k.json
+{"fixed_run_id_calls": 1, "benchmark_unchanged": true}

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/serving-analysis-01/REPORT.md

@@ -0,0 +1,19 @@
+# Actual cooled llm_decode_bench results
+
+| Test | Tokens/sec | Cooldown seconds |
+|---|---:|---:|
+| 32K prefill | 8457.000 | 92.3 |
+| 64K prefill | 8443.000 | 92.3 |
+| 128K prefill | 8323.000 | 92.3 |
+| 32K C1 decode | 204.930 | 92.3 |
+| 32K C4 decode aggregate | 329.343 | 92.3 |
+| 64K C1 decode | 199.446 | 92.3 |
+| 64K C4 decode aggregate | 330.096 | 92.4 |
+| 128K C1 decode | 204.445 | 92.3 |
+| 128K C4 decode aggregate | 334.086 | 92.3 |
+| 8K C8 decode aggregate | 594.639 | 92.3 |
+| 16K C8 decode aggregate | 602.965 | 92.3 |
+
+KV capacity: 23,562,091 tokens. NCCL8 channels verified from logs. Unchanged selected FP8 token-map-hoist image, TP4/DCP4/MTP3, graphs, NVFP4KV, maxseq24, batch4096, four300W GPUs.
+
+Each cell follows at least90seconds idle and all four GPUs at or below55C for30seconds. In-cell temperatures/clocks are retained. This reduces thermal differences but does not guarantee identical sustained clocks or output trajectories. This extends the prior shorter-context token-map-hoist run with unchanged image and runtime. New context cells have no matched control. This is a single-run exploratory comparison; output and MTP variation remain. Plain one-shot cutoff is131072bytes; fused cutoff86016bytes and two-shot maximum786432bytes are verified from runtime logs. No adoption or publication. Production and clients remain stopped.

results/speed-20260909/evidence/tile-policy-route-hoist-expanded-02/serving-analysis-01/SUMMARY.json

@@ -0,0 +1,366 @@
+{
+  "scope": "Expanded cooled FP8 route-hoist context and concurrency screen; no independent replication or model quality qualification",
+  "image": "sha256:ca6b80188dce154b91f49108b7d87792d2ba6328935afc71b44d1c0e6f6a1adf",
+  "kv_tokens": 23562091,
+  "channel_log_lines": [
+    "<host>:810:810 [0] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:812:812 [2] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:813:813 [3] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:811:811 [1] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:810:810 [0] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:811:811 [1] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:813:813 [3] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:812:812 [2] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:810:810 [0] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:811:811 [1] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:813:813 [3] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:812:812 [2] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:812:812 [2] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:813:813 [3] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:810:810 [0] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer",
+    "<host>:811:811 [1] NCCL INFO 8 coll channels, 8 collnet channels, 0 nvls channels, 8 p2p channels, 2 p2p channels per peer"
+  ],
+  "runs": [
+    {
+      "test": "32K prefill",
+      "tokens_per_second": 8457.0,
+      "cooldown_seconds": 92.26148488814943,
+      "pre_cell_hardware_csv": "0, 40, 172, 9.17, 300.00\n1, 31, 180, 14.84, 300.00\n2, 33, 180, 6.61, 300.00\n3, 33, 180, 17.29, 300.00\n",
+      "during_cell_hardware": {
+        "samples": 9,
+        "duration_seconds": 19.188,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.71,
+        "cpu_temp_max_c": 75.25,
+        "gpu_util_avg_pct": 95.08,
+        "gpu_util_max_pct": 100.0,
+        "mem_util_avg_pct": 27.33,
+        "mem_util_max_pct": 36.0,
+        "temp_avg_c": 48.33,
+        "temp_max_c": 62.0,
+        "power_total_avg_w": 1104.35,
+        "power_total_max_w": 1135.15,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 47818.11,
+        "pcie_rx_max_mb_s": 65384.0,
+        "pcie_tx_avg_mb_s": 41943.22,
+        "pcie_tx_max_mb_s": 54316.0
+      }
+    },
+    {
+      "test": "64K prefill",
+      "tokens_per_second": 8443.0,
+      "cooldown_seconds": 92.32916426914744,
+      "pre_cell_hardware_csv": "0, 42, 180, 9.15, 300.00\n1, 32, 180, 15.18, 300.00\n2, 34, 180, 6.41, 300.00\n3, 34, 180, 18.21, 300.00\n",
+      "during_cell_hardware": {
+        "samples": 10,
+        "duration_seconds": 21.613,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.76,
+        "cpu_temp_max_c": 74.12,
+        "gpu_util_avg_pct": 96.5,
+        "gpu_util_max_pct": 100.0,
+        "mem_util_avg_pct": 26.88,
+        "mem_util_max_pct": 35.0,
+        "temp_avg_c": 49.95,
+        "temp_max_c": 65.0,
+        "power_total_avg_w": 1095.48,
+        "power_total_max_w": 1129.79,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 51563.0,
+        "pcie_rx_max_mb_s": 60865.0,
+        "pcie_tx_avg_mb_s": 51753.4,
+        "pcie_tx_max_mb_s": 61757.0
+      }
+    },
+    {
+      "test": "128K prefill",
+      "tokens_per_second": 8323.0,
+      "cooldown_seconds": 92.31365287397057,
+      "pre_cell_hardware_csv": "0, 43, 180, 9.96, 300.00\n1, 33, 180, 15.40, 300.00\n2, 35, 180, 6.89, 300.00\n3, 34, 180, 18.67, 300.00\n",
+      "during_cell_hardware": {
+        "samples": 13,
+        "duration_seconds": 28.84,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.81,
+        "cpu_temp_max_c": 75.38,
+        "gpu_util_avg_pct": 99.98,
+        "gpu_util_max_pct": 100.0,
+        "mem_util_avg_pct": 27.08,
+        "mem_util_max_pct": 32.0,
+        "temp_avg_c": 53.13,
+        "temp_max_c": 68.0,
+        "power_total_avg_w": 1120.21,
+        "power_total_max_w": 1128.52,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 53419.31,
+        "pcie_rx_max_mb_s": 60487.0,
+        "pcie_tx_avg_mb_s": 51199.31,
+        "pcie_tx_max_mb_s": 58392.0
+      }
+    },
+    {
+      "test": "32K C1 decode",
+      "tokens_per_second": 204.93040297129508,
+      "cooldown_seconds": 92.28490831400268,
+      "pre_cell_hardware_csv": "0, 45, 180, 11.04, 300.00\n1, 33, 180, 15.35, 300.00\n2, 37, 180, 7.42, 300.00\n3, 36, 180, 18.46, 300.00\n",
+      "during_cell_hardware": {
+        "samples": 9,
+        "duration_seconds": 19.264,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.88,
+        "cpu_temp_max_c": 76.0,
+        "gpu_util_avg_pct": 99.0,
+        "gpu_util_max_pct": 99.0,
+        "mem_util_avg_pct": 55.67,
+        "mem_util_max_pct": 61.0,
+        "temp_avg_c": 55.0,
+        "temp_max_c": 70.0,
+        "power_total_avg_w": 1162.79,
+        "power_total_max_w": 1167.42,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 9085.0,
+        "pcie_rx_max_mb_s": 9176.0,
+        "pcie_tx_avg_mb_s": 8975.11,
+        "pcie_tx_max_mb_s": 9069.0
+      }
+    },
+    {
+      "test": "32K C4 decode aggregate",
+      "tokens_per_second": 329.34289874948905,
+      "cooldown_seconds": 92.34142689802684,
+      "pre_cell_hardware_csv": "0, 46, 180, 10.95, 300.00\n1, 34, 180, 15.30, 300.00\n2, 37, 180, 7.04, 300.00\n3, 36, 180, 19.14, 300.00\n",
+      "during_cell_hardware": {
+        "samples": 8,
+        "duration_seconds": 16.84,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.83,
+        "cpu_temp_max_c": 76.25,
+        "gpu_util_avg_pct": 100.0,
+        "gpu_util_max_pct": 100.0,
+        "mem_util_avg_pct": 43.84,
+        "mem_util_max_pct": 51.0,
+        "temp_avg_c": 56.12,
+        "temp_max_c": 71.0,
+        "power_total_avg_w": 1175.11,
+        "power_total_max_w": 1178.19,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 11227.62,
+        "pcie_rx_max_mb_s": 11525.0,
+        "pcie_tx_avg_mb_s": 11234.75,
+        "pcie_tx_max_mb_s": 11531.0
+      }
+    },
+    {
+      "test": "64K C1 decode",
+      "tokens_per_second": 199.4458582488372,
+      "cooldown_seconds": 92.34245188813657,
+      "pre_cell_hardware_csv": "0, 46, 180, 10.38, 300.00\n1, 34, 180, 15.44, 300.00\n2, 37, 180, 7.34, 300.00\n3, 36, 180, 19.35, 300.00\n",
+      "during_cell_hardware": {
+        "samples": 8,
+        "duration_seconds": 16.873,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.91,
+        "cpu_temp_max_c": 75.5,
+        "gpu_util_avg_pct": 99.0,
+        "gpu_util_max_pct": 99.0,
+        "mem_util_avg_pct": 55.41,
+        "mem_util_max_pct": 61.0,
+        "temp_avg_c": 57.12,
+        "temp_max_c": 72.0,
+        "power_total_avg_w": 1167.07,
+        "power_total_max_w": 1170.68,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 9075.25,
+        "pcie_rx_max_mb_s": 9156.0,
+        "pcie_tx_avg_mb_s": 8901.88,
+        "pcie_tx_max_mb_s": 8976.0
+      }
+    },
+    {
+      "test": "64K C4 decode aggregate",
+      "tokens_per_second": 330.0961767418383,
+      "cooldown_seconds": 92.35157263395377,
+      "pre_cell_hardware_csv": "0, 47, 180, 10.64, 300.00\n1, 34, 180, 15.21, 300.00\n2, 38, 180, 7.34, 300.00\n3, 37, 180, 19.59, 300.00\n",
+      "during_cell_hardware": {
+        "samples": 9,
+        "duration_seconds": 19.217,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.83,
+        "cpu_temp_max_c": 76.25,
+        "gpu_util_avg_pct": 100.0,
+        "gpu_util_max_pct": 100.0,
+        "mem_util_avg_pct": 44.33,
+        "mem_util_max_pct": 52.0,
+        "temp_avg_c": 56.03,
+        "temp_max_c": 72.0,
+        "power_total_avg_w": 1175.71,
+        "power_total_max_w": 1180.0,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 11146.11,
+        "pcie_rx_max_mb_s": 11547.0,
+        "pcie_tx_avg_mb_s": 11112.44,
+        "pcie_tx_max_mb_s": 11387.0
+      }
+    },
+    {
+      "test": "128K C1 decode",
+      "tokens_per_second": 204.44477860988547,
+      "cooldown_seconds": 92.32274625892751,
+      "pre_cell_hardware_csv": "0, 47, 180, 11.07, 300.00\n1, 34, 180, 15.44, 300.00\n2, 37, 180, 7.11, 300.00\n3, 37, 180, 19.24, 300.00\n",
+      "during_cell_hardware": {
+        "samples": 8,
+        "duration_seconds": 16.831,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.88,
+        "cpu_temp_max_c": 75.62,
+        "gpu_util_avg_pct": 99.0,
+        "gpu_util_max_pct": 99.0,
+        "mem_util_avg_pct": 55.0,
+        "mem_util_max_pct": 61.0,
+        "temp_avg_c": 59.72,
+        "temp_max_c": 75.0,
+        "power_total_avg_w": 1172.01,
+        "power_total_max_w": 1176.12,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 8998.38,
+        "pcie_rx_max_mb_s": 9099.0,
+        "pcie_tx_avg_mb_s": 8845.62,
+        "pcie_tx_max_mb_s": 8970.0
+      }
+    },
+    {
+      "test": "128K C4 decode aggregate",
+      "tokens_per_second": 334.08648489959006,
+      "cooldown_seconds": 92.34488186100498,
+      "pre_cell_hardware_csv": "0, 48, 180, 10.30, 300.00\n1, 35, 180, 15.50, 300.00\n2, 38, 180, 7.72, 300.00\n3, 38, 180, 20.00, 300.00\n",
+      "during_cell_hardware": {
+        "samples": 9,
+        "duration_seconds": 19.194,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.82,
+        "cpu_temp_max_c": 76.62,
+        "gpu_util_avg_pct": 100.0,
+        "gpu_util_max_pct": 100.0,
+        "mem_util_avg_pct": 43.08,
+        "mem_util_max_pct": 51.0,
+        "temp_avg_c": 56.31,
+        "temp_max_c": 72.0,
+        "power_total_avg_w": 1175.4,
+        "power_total_max_w": 1179.06,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 11116.0,
+        "pcie_rx_max_mb_s": 11670.0,
+        "pcie_tx_avg_mb_s": 11087.89,
+        "pcie_tx_max_mb_s": 11381.0
+      }
+    },
+    {
+      "test": "8K C8 decode aggregate",
+      "tokens_per_second": 594.6390215769761,
+      "cooldown_seconds": 92.34947287291288,
+      "pre_cell_hardware_csv": "0, 47, 180, 10.47, 300.00\n1, 34, 180, 15.13, 300.00\n2, 37, 180, 7.10, 300.00\n3, 37, 180, 21.71, 300.00\n",
+      "during_cell_hardware": {
+        "samples": 9,
+        "duration_seconds": 19.215,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.82,
+        "cpu_temp_max_c": 76.5,
+        "gpu_util_avg_pct": 100.0,
+        "gpu_util_max_pct": 100.0,
+        "mem_util_avg_pct": 51.17,
+        "mem_util_max_pct": 59.0,
+        "temp_avg_c": 56.5,
+        "temp_max_c": 72.0,
+        "power_total_avg_w": 1169.91,
+        "power_total_max_w": 1173.63,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 14848.56,
+        "pcie_rx_max_mb_s": 16432.0,
+        "pcie_tx_avg_mb_s": 14370.89,
+        "pcie_tx_max_mb_s": 15808.0
+      }
+    },
+    {
+      "test": "16K C8 decode aggregate",
+      "tokens_per_second": 602.9645987499702,
+      "cooldown_seconds": 92.3161594690755,
+      "pre_cell_hardware_csv": "0, 47, 180, 10.47, 300.00\n1, 34, 180, 15.41, 300.00\n2, 37, 180, 7.35, 300.00\n3, 37, 180, 19.90, 300.00\n",
+      "during_cell_hardware": {
+        "samples": 8,
+        "duration_seconds": 16.811,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.99,
+        "cpu_temp_max_c": 76.62,
+        "gpu_util_avg_pct": 100.0,
+        "gpu_util_max_pct": 100.0,
+        "mem_util_avg_pct": 51.25,
+        "mem_util_max_pct": 59.0,
+        "temp_avg_c": 57.06,
+        "temp_max_c": 72.0,
+        "power_total_avg_w": 1171.73,
+        "power_total_max_w": 1174.9,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384370.0,
+        "vram_used_max_mb": 384370.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.17,
+        "vram_used_max_pct": 98.17,
+        "pcie_rx_avg_mb_s": 15423.5,
+        "pcie_rx_max_mb_s": 17029.0,
+        "pcie_tx_avg_mb_s": 14292.0,
+        "pcie_tx_max_mb_s": 15343.0
+      }
+    }
+  ]
+}

results/speed-20260909/evidence/tile-policy-route-hoist-m1-component-01/CPU-RESULT.json

@@ -0,0 +1,7 @@
+{
+  "status": "passed",
+  "executed_candidate_helper_cases": 117760,
+  "token_mapping_comparisons": 268800,
+  "candidate_sha256": "5bad78a538d2f1ff3a0a6bd3a9c7ae640c56a4ea75fe5db5493f27972272cad1",
+  "scope": "CPU executes actual candidate cache helper AST and compares original token-map formula, allvalid tails, broadcast and deterministic modes. No GPU or speed claim."
+}

results/speed-20260909/evidence/tile-policy-route-hoist-m1-component-01/SUMMARY.json

@@ -0,0 +1,147 @@
+{
+  "exact_comparisons": 66,
+  "fp8_artifacts": 2,
+  "timings": [
+    {
+      "arm": "r0-k4-ordinary",
+      "m": 1,
+      "capture": 0,
+      "reference_median_ms": 0.0797648012638092,
+      "candidate_median_ms": 0.07784799933433532,
+      "paired_delta_median_ms": -0.0019520044326782338,
+      "samples": [
+        {
+          "sample": 0,
+          "baseline_ms": 0.08009759783744812,
+          "candidate_ms": 0.07788000106811524
+        },
+        {
+          "sample": 1,
+          "candidate_ms": 0.07785279750823974,
+          "baseline_ms": 0.07987679839134217
+        },
+        {
+          "sample": 2,
+          "baseline_ms": 0.0797648012638092,
+          "candidate_ms": 0.07784799933433532
+        },
+        {
+          "sample": 3,
+          "candidate_ms": 0.07780479788780212,
+          "baseline_ms": 0.07975680232048035
+        },
+        {
+          "sample": 4,
+          "baseline_ms": 0.07968639731407165,
+          "candidate_ms": 0.07781760096549988
+        }
+      ]
+    },
+    {
+      "arm": "r0-k4-ordinary",
+      "m": 1,
+      "capture": 1,
+      "reference_median_ms": 0.07989439964294434,
+      "candidate_median_ms": 0.07783520221710205,
+      "paired_delta_median_ms": -0.002041596174240115,
+      "samples": [
+        {
+          "sample": 0,
+          "baseline_ms": 0.08003039956092835,
+          "candidate_ms": 0.0778656005859375
+        },
+        {
+          "sample": 1,
+          "candidate_ms": 0.07781919836997986,
+          "baseline_ms": 0.07989439964294434
+        },
+        {
+          "sample": 2,
+          "baseline_ms": 0.07987679839134217,
+          "candidate_ms": 0.07783520221710205
+        },
+        {
+          "sample": 3,
+          "candidate_ms": 0.0778880000114441,
+          "baseline_ms": 0.07989760041236878
+        },
+        {
+          "sample": 4,
+          "baseline_ms": 0.07977439761161804,
+          "candidate_ms": 0.0778335988521576
+        }
+      ]
+    },
+    {
+      "arm": "r0-k5-ordinary",
+      "m": 1,
+      "capture": 0,
+      "reference_median_ms": 0.08399999737739564,
+      "candidate_median_ms": 0.08184639811515808,
+      "paired_delta_median_ms": -0.002187198400497428,
+      "samples": [
+        {
+          "sample": 0,
+          "baseline_ms": 0.08425120115280152,
+          "candidate_ms": 0.08194400072097778
+        },
+        {
+          "sample": 1,
+          "candidate_ms": 0.0819487988948822,
+          "baseline_ms": 0.08407520055770874
+        },
+        {
+          "sample": 2,
+          "baseline_ms": 0.08398560285568238,
+          "candidate_ms": 0.08184639811515808
+        },
+        {
+          "sample": 3,
+          "candidate_ms": 0.08161439895629882,
+          "baseline_ms": 0.08399999737739564
+        },
+        {
+          "sample": 4,
+          "baseline_ms": 0.08399199843406677,
+          "candidate_ms": 0.08180480003356934
+        }
+      ]
+    },
+    {
+      "arm": "r0-k5-ordinary",
+      "m": 1,
+      "capture": 1,
+      "reference_median_ms": 0.08403199911117554,
+      "candidate_median_ms": 0.08174719810485839,
+      "paired_delta_median_ms": -0.0023360013961792048,
+      "samples": [
+        {
+          "sample": 0,
+          "baseline_ms": 0.08424479961395263,
+          "candidate_ms": 0.08177760243415833
+        },
+        {
+          "sample": 1,
+          "candidate_ms": 0.0815280020236969,
+          "baseline_ms": 0.0839792013168335
+        },
+        {
+          "sample": 2,
+          "baseline_ms": 0.0840831995010376,
+          "candidate_ms": 0.08174719810485839
+        },
+        {
+          "sample": 3,
+          "candidate_ms": 0.08171520233154297,
+          "baseline_ms": 0.0839904010295868
+        },
+        {
+          "sample": 4,
+          "baseline_ms": 0.08403199911117554,
+          "candidate_ms": 0.08184639811515808
+        }
+      ]
+    }
+  ],
+  "scope": "Representative full-local-MoE component milliseconds only; not llm_decode_bench tokens/sec or quality qualification. M1 coverage supplements the M4/M16/M4096 matrix."
+}

results/speed-20260909/evidence/tile-policy-route-hoist-m1-component-01/results-01/result.json

@@ -0,0 +1,28 @@
+{
+  "status": "component-matrix-completed",
+  "results": [
+    [
+      {
+        "status": "passed",
+        "comparison_count": 33,
+        "raw_sha256": "6caa73e2e319fdfc71dce0924ce736621e942c16a79dbc63973b58cf451f767e",
+        "mode": "ordinary",
+        "rank": 0,
+        "bits": 4,
+        "scope": "representative component checks; not throughput or model quality"
+      },
+      {
+        "status": "passed",
+        "comparison_count": 33,
+        "raw_sha256": "3a2e8795e11176b7f525f813f893511d4fe0c618654d549c9ada4ec6a852ea98",
+        "mode": "ordinary",
+        "rank": 0,
+        "bits": 5,
+        "scope": "representative component checks; not throughput or model quality"
+      }
+    ]
+  ],
+  "resource_audit": "pending; retained fresh compile caches",
+  "throughput": "Not measured",
+  "quality": "Not qualified"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/result.json

@@ -0,0 +1,4 @@
+{
+  "status": "tile-policy-route-hoist-speed-screen-completed",
+  "quality": "qualification pending"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c1-command.json

@@ -0,0 +1,29 @@
+[
+  "/usr/bin/python3",
+  "<campaign>/fixed-workload-candidate-comparison-01/bench_fixed.py",
+  "--host",
+  "127.0.0.1",
+  "--port",
+  "8001",
+  "--model",
+  "glm53-flash-trellismx-p8-k45",
+  "--duration",
+  "20",
+  "--max-tokens",
+  "8192",
+  "--token-targeting",
+  "exact",
+  "--display-mode",
+  "plain",
+  "--output",
+  "<campaign>/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c1.json",
+  "--contexts",
+  "8k",
+  "--concurrency",
+  "1",
+  "--skip-prefill",
+  "--cell-warmup-timeout-seconds",
+  "180",
+  "--temperature",
+  "0"
+]

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c1-cooldown.json

@@ -0,0 +1,84 @@
+{
+  "status": "passed",
+  "minimum_seconds": 90,
+  "temperature_ceiling_c": 55,
+  "stable_seconds": 30,
+  "samples": [
+    {
+      "elapsed_seconds": 2.600718289613724e-07,
+      "hardware_csv": "0, 59, 2310, 106.96, 300.00\n1, 43, 2872, 140.71, 300.00\n2, 50, 2310, 84.04, 300.00\n3, 45, 2850, 138.24, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 5.097404036205262,
+      "hardware_csv": "0, 57, 2272, 76.53, 300.00\n1, 41, 2610, 86.56, 300.00\n2, 48, 2272, 74.03, 300.00\n3, 44, 2610, 98.62, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 10.194114323006943,
+      "hardware_csv": "0, 56, 2272, 75.02, 300.00\n1, 41, 2610, 86.27, 300.00\n2, 47, 2272, 73.31, 300.00\n3, 43, 2610, 98.17, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 15.304537658113986,
+      "hardware_csv": "0, 56, 2272, 73.50, 300.00\n1, 40, 2610, 86.06, 300.00\n2, 47, 2272, 72.89, 300.00\n3, 43, 2610, 98.16, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 20.41973445098847,
+      "hardware_csv": "0, 53, 180, 14.09, 300.00\n1, 39, 180, 29.26, 300.00\n2, 44, 180, 10.97, 300.00\n3, 41, 180, 33.43, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 25.560495782177895,
+      "hardware_csv": "0, 52, 180, 13.01, 300.00\n1, 38, 180, 16.28, 300.00\n2, 43, 180, 9.94, 300.00\n3, 40, 180, 22.44, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 30.69500490813516,
+      "hardware_csv": "0, 51, 180, 12.64, 300.00\n1, 38, 180, 15.54, 300.00\n2, 42, 180, 9.50, 300.00\n3, 40, 180, 21.81, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 35.82530016917735,
+      "hardware_csv": "0, 50, 180, 12.28, 300.00\n1, 37, 180, 16.94, 300.00\n2, 42, 180, 9.37, 300.00\n3, 39, 180, 22.11, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 40.956741880159825,
+      "hardware_csv": "0, 49, 180, 11.84, 300.00\n1, 37, 180, 15.53, 300.00\n2, 41, 180, 8.97, 300.00\n3, 39, 180, 21.52, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 46.098311140201986,
+      "hardware_csv": "0, 49, 180, 11.15, 300.00\n1, 36, 180, 15.49, 300.00\n2, 40, 180, 8.41, 300.00\n3, 38, 180, 21.71, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 51.231507680146024,
+      "hardware_csv": "0, 48, 180, 11.57, 300.00\n1, 36, 180, 15.48, 300.00\n2, 40, 180, 8.36, 300.00\n3, 38, 180, 21.63, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 56.367206645198166,
+      "hardware_csv": "0, 48, 180, 10.57, 300.00\n1, 35, 180, 15.33, 300.00\n2, 39, 180, 7.62, 300.00\n3, 38, 180, 21.29, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 61.499323948984966,
+      "hardware_csv": "0, 47, 180, 10.38, 300.00\n1, 35, 180, 15.43, 300.00\n2, 39, 180, 7.74, 300.00\n3, 37, 180, 20.87, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 66.62913945014589,
+      "hardware_csv": "0, 47, 180, 10.19, 300.00\n1, 35, 180, 15.53, 300.00\n2, 38, 180, 7.07, 300.00\n3, 37, 180, 21.27, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 71.76942849415354,
+      "hardware_csv": "0, 46, 180, 10.40, 300.00\n1, 35, 180, 15.45, 300.00\n2, 38, 180, 7.69, 300.00\n3, 37, 180, 21.06, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 76.91019542212598,
+      "hardware_csv": "0, 46, 180, 10.12, 300.00\n1, 34, 180, 15.39, 300.00\n2, 37, 180, 7.48, 300.00\n3, 36, 180, 20.29, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 82.05304709100164,
+      "hardware_csv": "0, 45, 180, 9.99, 300.00\n1, 34, 180, 15.26, 300.00\n2, 37, 180, 7.50, 300.00\n3, 36, 180, 19.62, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 87.19396727904677,
+      "hardware_csv": "0, 45, 180, 9.80, 300.00\n1, 34, 180, 15.29, 300.00\n2, 36, 180, 7.39, 300.00\n3, 36, 180, 20.24, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 92.32664809399284,
+      "hardware_csv": "0, 44, 180, 9.83, 300.00\n1, 33, 180, 15.25, 300.00\n2, 36, 180, 7.17, 300.00\n3, 36, 180, 20.03, 300.00\n"
+    }
+  ]
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c1-fixed-workload.json

@@ -0,0 +1,5 @@
+{
+  "prefix": "tmxrepeatabc",
+  "temperature": 0,
+  "scope": "greedy fixed-prompt diagnostic"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c1-receipt.json

@@ -0,0 +1,5 @@
+{
+  "exit_code": 0,
+  "result_exists": true,
+  "sha256": "63f1d4cca98af9695b8f03ec470a7a8d5311843c10cd0a81f76703720b2177cb"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c1.json

@@ -0,0 +1,345 @@
+{
+  "metadata": {
+    "version": "0.4.29",
+    "engine": "vllm",
+    "model": "glm53-flash-trellismx-p8-k45",
+    "server": "127.0.0.1:8001",
+    "timestamp": "2026-09-09T13:56:34.071188",
+    "decode_mode": "duration",
+    "primary_decode_layer": "sustained_decode",
+    "duration_per_test": 20.0,
+    "request_count": 0,
+    "warmup_request_count": 0,
+    "run_burst": false,
+    "prefill_mode": "skipped",
+    "standalone_prefill": false,
+    "prefill_only": false,
+    "skip_prefill": true,
+    "burst_e2e_status": "not_run_use_--run-burst",
+    "burst_request_count": 0,
+    "burst_warmup_request_count": 0,
+    "burst_requests_per_concurrency": 5,
+    "decode_warmup_seconds": 3.0,
+    "decode_warmup_context": 8192,
+    "decode_warmup_concurrency": 1,
+    "cell_warmup_timeout_seconds": 180.0,
+    "cell_warmup_timeout_policy": "<=32k:60s,64k:120s,>=128k:180s when override is 0",
+    "show_capacity_limited_values": false,
+    "max_tokens": 8192,
+    "temperature": 0.0,
+    "ignore_eos": true,
+    "max_total_tokens": 29540352,
+    "dcp_size": 0,
+    "metrics_available": true,
+    "metrics_warning": "",
+    "concurrency_levels": [
+      1
+    ],
+    "context_lengths": [
+      8192
+    ],
+    "startup_diagnostics_available": true,
+    "nvidia_p2p_override_effective": true,
+    "p2pmark_status": "not_run",
+    "amd_fabric_status": "not_run"
+  },
+  "startup_diagnostics": {
+    "version": "0.4.29",
+    "server_url": "http://127.0.0.1:8001",
+    "hostname": "<host>",
+    "uname": "Linux <host> 6.18.7-76061807-generic #202601231045~1769703228~24.04~cb87b5b SMP PREEMPT_DYNAMIC Thu J x86_64 x86_64 x86_64 GNU/Linux",
+    "env": {},
+    "args": {
+      "concurrency": "1",
+      "contexts": "8k",
+      "max_tokens": 8192,
+      "duration": 20.0,
+      "request_count": 0,
+      "run_burst": false,
+      "standalone_prefill": false,
+      "prefill_only": false,
+      "skip_prefill": true,
+      "prefill_contexts": "8k,64k,128k",
+      "prefill_metric": "client",
+      "dcp_size": 0,
+      "kv_budget": 0
+    },
+    "nvidia_p2p_override": {
+      "effective": true,
+      "configured": true,
+      "params_path": "/proc/driver/nvidia/params",
+      "params_available": true,
+      "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+      "modprobe_available": true,
+      "runtime": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1",
+        "DmaRemapPeerMmio": "1"
+      },
+      "expected": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1"
+      },
+      "missing": [],
+      "mismatched": {},
+      "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+      "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+      "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+    },
+    "p2pmark": {
+      "status": "not_run"
+    },
+    "amd_fabric": {
+      "status": "not_run"
+    },
+    "nvidia_smi_query": {
+      "cmd": [
+        "nvidia-smi",
+        "--query-gpu=index,name,driver_version,pci.bus_id,pcie.link.gen.current,pcie.link.width.current,power.limit",
+        "--format=csv,noheader,nounits"
+      ],
+      "returncode": 0,
+      "stdout": "0, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:01:00.0, 1, 16, 300.00\n1, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:21:00.0, 1, 16, 300.00\n2, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:81:00.0, 1, 16, 300.00\n3, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:C1:00.0, 1, 16, 300.00",
+      "stderr": ""
+    },
+    "nvidia_smi_topo": {
+      "cmd": [
+        "nvidia-smi",
+        "topo",
+        "-m"
+      ],
+      "returncode": 0,
+      "stdout": "\u001b[4mGPU0\tGPU1\tGPU2\tGPU3\tCPU Affinity\tNUMA Affinity\tGPU NUMA ID\u001b[0m\nGPU0\t X \tNODE\tNODE\tNODE\t0-47\t0\t\tN/A\nGPU1\tNODE\t X \tNODE\tNODE\t0-47\t0\t\tN/A\nGPU2\tNODE\tNODE\t X \tNODE\t0-47\t0\t\tN/A\nGPU3\tNODE\tNODE\tNODE\t X \t0-47\t0\t\tN/A\n\nLegend:\n\n  X    = Self\n  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)\n  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node\n  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)\n  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)\n  PIX  = Connection traversing at most a single PCIe bridge\n  NV#  = Connection traversing a bonded set of # NVLinks",
+      "stderr": ""
+    }
+  },
+  "nvidia_p2p_override": {
+    "effective": true,
+    "configured": true,
+    "params_path": "/proc/driver/nvidia/params",
+    "params_available": true,
+    "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+    "modprobe_available": true,
+    "runtime": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1",
+      "DmaRemapPeerMmio": "1"
+    },
+    "expected": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1"
+    },
+    "missing": [],
+    "mismatched": {},
+    "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+    "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+    "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+  },
+  "p2pmark": {
+    "status": "not_run"
+  },
+  "amd_fabric": {
+    "status": "not_run"
+  },
+  "hardware_run_summary": {
+    "samples": 16,
+    "duration_seconds": 36.083,
+    "gpu_count": 4,
+    "cpu_util_avg_pct": 9.8,
+    "cpu_temp_max_c": 76.62,
+    "gpu_util_avg_pct": 86.36,
+    "gpu_util_max_pct": 100.0,
+    "mem_util_avg_pct": 45.42,
+    "mem_util_max_pct": 62.0,
+    "temp_avg_c": 50.03,
+    "temp_max_c": 68.0,
+    "power_total_avg_w": 1017.7,
+    "power_total_max_w": 1162.92,
+    "power_limit_total_w": 1200.0,
+    "vram_used_avg_mb": 384450.0,
+    "vram_used_max_mb": 384450.0,
+    "vram_total_mb": 391548.0,
+    "vram_used_avg_pct": 98.19,
+    "vram_used_max_pct": 98.19,
+    "pcie_rx_avg_mb_s": 8000.75,
+    "pcie_rx_max_mb_s": 9534.0,
+    "pcie_tx_avg_mb_s": 7876.0,
+    "pcie_tx_max_mb_s": 9256.0
+  },
+  "event_log": [
+    "13:55:56 benchmark start engine=vllm",
+    "13:55:56 startup server=http://127.0.0.1:8001 model=glm53-flash-trellismx-p8-k45",
+    "13:55:56 startup decode concurrency=1 contexts=8k",
+    "13:55:56 startup NVIDIA P2P override: enabled: runtime NVIDIA P2P override matches expected RegistryDwords",
+    "13:55:56 startup engine vLLM 0.26.1rc0+glm53.flash.nvfp4.luke.clean.r1.vllme75bcfd.b12x58a046f models=['glm53-flash-trellismx-p8-k45']",
+    "13:55:56 startup KV cache budget from vLLM metrics: 29,540,352 tokens (3606 blocks x 2048; local 7,385,088 \u00d7 CP 4; CP source: local process)",
+    "13:55:56 startup model context length: 1,000,000 tokens",
+    "13:55:56 startup prefill tests: skipped",
+    "13:55:56 startup calibrating padding text run=tmxrepeatabc up_to=8k",
+    "13:55:56 startup context 8k: 50,558 chars (8,192 prompt tokens via /tokenize)",
+    "13:55:56 startup token targeting: /tokenize exact",
+    "13:55:56 startup startup preparation done",
+    "13:55:56 hardware monitor interval=2s",
+    "13:55:56 decode warmup start",
+    "13:55:56 decode warmup start C=1 ctx=8k 3s",
+    "13:55:56 cell start C=1 ctx=8k",
+    "13:56:01 ready C=1 ctx=8k running_reqs=1/1, queue_reqs=0, active_streams=1/1, stable=3.0s",
+    "13:56:04 cell done C=1 ctx=8k 207.0 tok/s",
+    "13:56:04 decode warmup done C=1 ctx=8k",
+    "13:56:06 cell start C=1 ctx=8k",
+    "13:56:12 ready C=1 ctx=8k running_reqs=1/1, queue_reqs=0, active_streams=1/1, stable=3.0s",
+    "13:56:32 cell done C=1 ctx=8k 222.1 tok/s"
+  ],
+  "prefill": {},
+  "results": [
+    {
+      "concurrency": 1,
+      "context_tokens": 8192,
+      "benchmark_mode": "duration",
+      "request_count_target": 0,
+      "warmup_request_count": 0,
+      "measurement_seconds": 19.991015,
+      "measurement_wall_seconds": 20.001145,
+      "client_output_tokens": 4440,
+      "server_output_tokens": 4440,
+      "aggregate_source": "openai_continuous_usage",
+      "aggregate_tps": 222.09977568136074,
+      "per_request_avg_tps": 222.09977568136074,
+      "ttft_avg": 0.5271740630269051,
+      "ttft_p50": 0.5271740630269051,
+      "ttft_p90": 0.5271740630269051,
+      "ttft_p99": 0.5271740630269051,
+      "time_to_second_token_avg": 0.015037497039884329,
+      "time_to_second_token_p50": 0.015037497039884329,
+      "time_to_second_token_p90": 0.015037497039884329,
+      "time_to_second_token_p99": 0.015037497039884329,
+      "request_latency_avg": 0.0,
+      "request_latency_p50": 0.0,
+      "request_latency_p90": 0.0,
+      "request_latency_p99": 0.0,
+      "inter_token_latency_avg": 0.004488577405983971,
+      "inter_token_latency_p50": 0.004488577405983971,
+      "inter_token_latency_p90": 0.004488577405983971,
+      "inter_token_latency_p99": 0.004488577405983971,
+      "output_tps_per_user_avg": 222.78773641440262,
+      "output_tps_per_user_p50": 222.78773641440262,
+      "output_tps_per_user_p90": 222.78773641440262,
+      "output_tps_per_user_p99": 222.78773641440262,
+      "e2e_output_tps_per_user_avg": 0.0,
+      "e2e_output_tps_per_user_p50": 0.0,
+      "e2e_output_tps_per_user_p90": 0.0,
+      "e2e_output_tps_per_user_p99": 0.0,
+      "chunk_inter_token_latency_avg": 0.013053017538076346,
+      "chunk_inter_token_latency_p50": 0.013053017538076346,
+      "chunk_inter_token_latency_p90": 0.013053017538076346,
+      "chunk_inter_token_latency_p99": 0.013053017538076346,
+      "input_seq_len_avg": 8192.0,
+      "output_seq_len_avg": 5346.0,
+      "output_seq_len_p50": 5346.0,
+      "output_seq_len_p90": 5346.0,
+      "output_seq_len_p99": 5346.0,
+      "request_count": 1,
+      "completed_request_count": 0,
+      "request_samples": [
+        {
+          "ttft": 0.5271740630269051,
+          "time_to_second_token": 0.015037497039884329,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.004488577405983971,
+          "chunk_inter_token_latency_avg": 0.013053017538076346,
+          "input_tokens": 8192,
+          "output_tokens": 5346,
+          "output_tps_per_user": 222.78773641440262,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        }
+      ],
+      "total_tokens": 4440,
+      "wall_time": 25.55005850503221,
+      "num_completed": 1,
+      "num_errors": 0,
+      "server_gen_throughput": 221.93211692306912,
+      "server_utilization": 0.006102635228848796,
+      "server_spec_accept_rate": 0.6666666666666666,
+      "server_spec_accept_length": 0.0,
+      "avg_running_reqs": 1,
+      "max_running_reqs": 1,
+      "effective_concurrency": 1,
+      "avg_queue_reqs": 0,
+      "max_queue_reqs": 0,
+      "queue_fraction": 0.0,
+      "underfilled": false,
+      "warmup_timed_out": false,
+      "warmup_duration": 5.543,
+      "ready_reason": "running_reqs=1/1, queue_reqs=0, active_streams=1/1, stable=3.0s",
+      "timeout_reason": "",
+      "capacity_limited": false,
+      "hardware_summary": {
+        "samples": 8,
+        "duration_seconds": 16.884,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.9,
+        "cpu_temp_max_c": 76.62,
+        "gpu_util_avg_pct": 99.0,
+        "gpu_util_max_pct": 99.0,
+        "mem_util_avg_pct": 54.69,
+        "mem_util_max_pct": 62.0,
+        "temp_avg_c": 53.38,
+        "temp_max_c": 68.0,
+        "power_total_avg_w": 1157.77,
+        "power_total_max_w": 1162.92,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384450.0,
+        "vram_used_max_mb": 384450.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.19,
+        "vram_used_max_pct": 98.19,
+        "pcie_rx_avg_mb_s": 9073.0,
+        "pcie_rx_max_mb_s": 9137.0,
+        "pcie_tx_avg_mb_s": 8941.38,
+        "pcie_tx_max_mb_s": 9046.0
+      }
+    }
+  ],
+  "summary_table": {
+    "8192": {
+      "1": 222.09977568136074
+    }
+  },
+  "burst_results": [],
+  "burst_summary_table": {},
+  "methodology": {
+    "prefill": {
+      "name": "Prefill",
+      "present": false,
+      "mode": "skipped",
+      "formula": "prompt_tokens / TTFT",
+      "notes": "Default mode records the required decode scout request for each non-zero decode context, so normal runs do not pay for a separate prefill phase. Standalone mode repeats cold-prefill samples. Prometheus prefill counters, when available and uncontaminated, are stored as validation."
+    },
+    "sustained_decode": {
+      "name": "Sustained Decode",
+      "present": true,
+      "formula": "OpenAI stream usage completion_tokens per measured window; client chunk fallback only when continuous usage is unavailable",
+      "notes": "Duration-based steady-state cell after warmup. This is the main tuning/regression signal for kernels, NCCL, DCP, MTP, and scheduling. Prometheus metrics are stored as validation and scheduler state, not the default headline."
+    },
+    "burst_e2e_decode": {
+      "name": "Burst / E2E Decode",
+      "present": false,
+      "status": "not run; use --run-burst",
+      "formula": "sum(completion_tokens) / profiling_wall_time",
+      "notes": "Finite client-facing request burst using OpenAI stream usage. It includes request admission, scheduling, prefill/cache behavior, and completion."
+    }
+  }
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c1.log

@@ -0,0 +1,109 @@
+{"diagnostic_wrapper": "fixed-prefix-greedy", "benchmark_sha256": "7239958032ec10781d7db3efca23a7602bd9aeb94455a723e2634ab7d6fe546a", "run_id": "tmxrepeatabc", "temperature": 0.0, "scope": "Separate diagnostic, not model-default throughput"}
+
+New version available: v0.6.2 (current: v0.4.29)
+Upgrade and restart? [Y/n]: Skipping update.
+
+╭──────────────────────────── NVIDIA P2P Override ─────────────────────────────╮
+│ Effective: yes                                                               │
+│ Configured file: yes (/etc/modprobe.d/nvidia-p2p-override.conf)              │
+│ Runtime: ForceP2P=0x11; RMForceP2PType=1; RMPcieP2PType=2;                   │
+│ GrdmaPciTopoCheckOverride=1; EnableResizableBar=1; DmaRemapPeerMmio=1        │
+╰──────────────────────────────────────────────────────────────────────────────╯
+╭─────────────────────────────── Configuration ────────────────────────────────╮
+│ LLM Inference Benchmark                                                      │
+│ Model: glm53-flash-trellismx-p8-k45 @ 127.0.0.1:8001                         │
+│ Decode concurrency: [1]                                                      │
+│ Decode contexts: ['8k']                                                      │
+│ Duration: 20.0s per decode test | Max tokens: 8192                           │
+│ Pre-decode warmup: C=1 max-runnable context for 3s                           │
+│ Prefill: skipped | Sustained decode: 1 cells                                 │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Engine: vLLM 0.26.1rc0+glm53.flash.nvfp4.luke.clean.r1.vllme75bcfd.b12x58a046f  
+Models: ['glm53-flash-trellismx-p8-k45']
+KV cache budget (vLLM metrics): 29,540,352 tokens (3606 blocks × 2048; local 
+7,385,088 × CP 4; CP source: local process)
+Model context length: 1,000,000 tokens
+Prefill tests: skipped
+Calibrating padding text (run=tmxrepeatabc, up to 8k)...
+  8k: 50,558 chars (8,192 prompt tokens via /tokenize)
+  Token targeting: /tokenize exact
+Done.
+
+
+
+llm-decode-bench v0.4.29
+╭────────────────────────────────── Phase 2 ───────────────────────────────────╮
+│ Sustained Decode                                                             │
+│ Steady-state decode throughput after the engine has admitted the requested   │
+│ concurrency and passed warmup. Use this as the main tuning/regression signal │
+│ for kernels, NCCL, DCP, MTP, and scheduler changes.                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Aggregate tok/s  + TTFT/ITL 
+╭────────────┬─────────────╮
+│ ctx \ conc │           1 │
+├────────────┼─────────────┤
+│ 8k         │ 222.1 527/4 │
+╰────────────┴─────────────╯
+Sustained Decode: aggregate tok/s uses OpenAI stream usage by default 
+(continuous completion_tokens when the server supports it). Prometheus is kept 
+as validation/scheduler data.
+Aggregate source(s): openai_continuous_usage
+Per-Request tok/s     
+╭────────────┬───────╮
+│ ctx \ conc │     1 │
+├────────────┼───────┤
+│ 8k         │ 222.1 │
+╰────────────┴───────╯
+Client request      
+latency: p50 / p90  
+ms                  
+╭────────────┬─────╮
+│ ctx \ conc │   1 │
+├────────────┼─────┤
+│ 8k         │ —/— │
+╰────────────┴─────╯
+Aggregate cells show dim detail as TTFT ms / ITL ms for the same ctx/conc 
+coordinate. ITL is computed from observed generated tokens, including streams 
+stopped at the measurement boundary; a missing ITL means no stream produced at 
+least two measured output tokens. Per-request tok/s and request latency are 
+shown in separate per-cell matrices. Completion/sample counts and full 
+request-level distributions remain in JSON under request_samples.
+Sustained mode: client latency metrics explain request UX variance; aggregate 
+tok/s remains the primary throughput signal. 
+ITL=(last_token_time-first_token_time)/(output_tokens-1), user tok/s=1/ITL.
+Hardware Summary                                                                
+╭───┬─┬───────┬───────────┬───────┬─────────┬─────┬──────┬─────┬───────────────╮
+│ … │ │ mode  │ GPU avg/… │ Mem … │ W avg/… │ T … │ CPU… │ VR… │ PCIe rx/tx a… │
+├───┼─┼───────┼───────────┼───────┼─────────┼─────┼──────┼─────┼───────────────┤
+│ … │ │ sust… │    99/99% │   55% │ 1158/1… │ 68C │  77C │ 98… │     9073/8941 │
+╰───┴─┴───────┴───────────┴───────┴─────────┴─────┴──────┴─────┴───────────────╯
+╭─────────────────────── Whole-run GPU Power ───────────────────────╮
+│ avg 1,018 W | max 1,163 W | limit 1,200 W | over 36s | 16 samples │
+╰───────────────────────────────────────────────────────────────────╯
+Hardware summary is sampled from nvidia-smi during the measured part of each 
+cell. Whole-run GPU power is the sampled sum of GPU power draw across the 
+complete benchmark run, not wall-outlet system power. PCIe rx/tx is MB/s and is 
+a coarse live diagnostic, not a per-kernel NCCL profiler.
+
+╭────────────────────────────────── Phase 3 ───────────────────────────────────╮
+│ Burst / E2E Decode                                                           │
+│ Not run. Re-run with --run-burst to append a finite client-facing request    │
+│ burst after Sustained Decode. This is intentionally disabled by default      │
+│ because it adds another full decode matrix.                                  │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭────────────────────────────── Primary Summary ───────────────────────────────╮
+│ Primary matrices repeated last so the important numbers are visible without  │
+│ scrolling back through diagnostics.                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Aggregate decode tok/s
+╭────────────┬───────╮
+│ ctx \ conc │     1 │
+├────────────┼───────┤
+│ 8k         │ 222.1 │
+╰────────────┴───────╯
+
+Results saved to 
+<campaign>/tile-policy-route-
+hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c1.json
+{"fixed_run_id_calls": 1, "benchmark_unchanged": true}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c4-command.json

@@ -0,0 +1,29 @@
+[
+  "/usr/bin/python3",
+  "<campaign>/fixed-workload-candidate-comparison-01/bench_fixed.py",
+  "--host",
+  "127.0.0.1",
+  "--port",
+  "8001",
+  "--model",
+  "glm53-flash-trellismx-p8-k45",
+  "--duration",
+  "20",
+  "--max-tokens",
+  "8192",
+  "--token-targeting",
+  "exact",
+  "--display-mode",
+  "plain",
+  "--output",
+  "<campaign>/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c4.json",
+  "--contexts",
+  "8k",
+  "--concurrency",
+  "4",
+  "--skip-prefill",
+  "--cell-warmup-timeout-seconds",
+  "180",
+  "--temperature",
+  "0"
+]

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c4-cooldown.json

@@ -0,0 +1,84 @@
+{
+  "status": "passed",
+  "minimum_seconds": 90,
+  "temperature_ceiling_c": 55,
+  "stable_seconds": 30,
+  "samples": [
+    {
+      "elapsed_seconds": 4.6100467443466187e-07,
+      "hardware_csv": "0, 63, 2310, 117.15, 300.00\n1, 43, 2745, 111.14, 300.00\n2, 52, 2287, 77.80, 300.00\n3, 47, 2850, 127.28, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 5.101989975897595,
+      "hardware_csv": "0, 60, 2272, 80.71, 300.00\n1, 42, 2610, 87.22, 300.00\n2, 51, 2272, 75.66, 300.00\n3, 45, 2610, 99.29, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 10.212053410010412,
+      "hardware_csv": "0, 59, 2272, 78.17, 300.00\n1, 42, 2610, 86.94, 300.00\n2, 50, 2272, 74.57, 300.00\n3, 45, 2610, 99.10, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 15.317986412905157,
+      "hardware_csv": "0, 58, 2272, 77.13, 300.00\n1, 41, 480, 74.50, 300.00\n2, 48, 427, 66.49, 300.00\n3, 43, 397, 71.09, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 20.4341272439342,
+      "hardware_csv": "0, 56, 180, 14.73, 300.00\n1, 40, 180, 16.40, 300.00\n2, 46, 180, 10.64, 300.00\n3, 42, 180, 26.27, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 25.576139567885548,
+      "hardware_csv": "0, 54, 180, 14.22, 300.00\n1, 39, 180, 15.89, 300.00\n2, 45, 180, 10.60, 300.00\n3, 42, 180, 22.94, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 30.70582295488566,
+      "hardware_csv": "0, 53, 180, 13.93, 300.00\n1, 38, 180, 15.74, 300.00\n2, 44, 180, 10.24, 300.00\n3, 41, 180, 23.26, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 35.84495552093722,
+      "hardware_csv": "0, 52, 180, 13.01, 300.00\n1, 38, 180, 15.76, 300.00\n2, 43, 172, 9.36, 300.00\n3, 41, 180, 22.07, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 40.97766937711276,
+      "hardware_csv": "0, 52, 180, 12.91, 300.00\n1, 37, 180, 15.74, 300.00\n2, 43, 180, 9.37, 300.00\n3, 40, 180, 21.79, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 46.1139068659395,
+      "hardware_csv": "0, 51, 180, 12.45, 300.00\n1, 37, 180, 15.81, 300.00\n2, 42, 180, 9.13, 300.00\n3, 40, 180, 22.13, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 51.250628226902336,
+      "hardware_csv": "0, 50, 180, 12.73, 300.00\n1, 37, 180, 15.40, 300.00\n2, 41, 180, 8.60, 300.00\n3, 39, 180, 22.41, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 56.38842057297006,
+      "hardware_csv": "0, 49, 180, 11.47, 300.00\n1, 36, 180, 15.44, 300.00\n2, 40, 172, 8.34, 300.00\n3, 39, 180, 22.14, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 61.528941567987204,
+      "hardware_csv": "0, 49, 180, 11.59, 300.00\n1, 36, 180, 15.41, 300.00\n2, 40, 180, 8.34, 300.00\n3, 38, 180, 20.63, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 66.66934219910763,
+      "hardware_csv": "0, 48, 180, 11.09, 300.00\n1, 35, 180, 15.44, 300.00\n2, 39, 180, 8.02, 300.00\n3, 38, 180, 21.10, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 71.80853762291372,
+      "hardware_csv": "0, 48, 180, 10.69, 300.00\n1, 35, 180, 15.37, 300.00\n2, 39, 180, 7.87, 300.00\n3, 38, 180, 20.78, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 76.94687540992163,
+      "hardware_csv": "0, 47, 180, 10.44, 300.00\n1, 35, 180, 15.49, 300.00\n2, 38, 180, 7.70, 300.00\n3, 37, 180, 21.17, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 82.08932351204567,
+      "hardware_csv": "0, 47, 180, 10.59, 300.00\n1, 35, 180, 15.74, 300.00\n2, 38, 180, 7.47, 300.00\n3, 37, 180, 20.56, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 87.2239846859593,
+      "hardware_csv": "0, 46, 180, 10.23, 300.00\n1, 34, 172, 15.57, 300.00\n2, 37, 180, 7.49, 300.00\n3, 37, 180, 20.02, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 92.3643947339151,
+      "hardware_csv": "0, 46, 180, 10.20, 300.00\n1, 34, 180, 15.34, 300.00\n2, 37, 180, 7.50, 300.00\n3, 36, 180, 20.28, 300.00\n"
+    }
+  ]
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c4-fixed-workload.json

@@ -0,0 +1,5 @@
+{
+  "prefix": "tmxrepeatabc",
+  "temperature": 0,
+  "scope": "greedy fixed-prompt diagnostic"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c4-receipt.json

@@ -0,0 +1,5 @@
+{
+  "exit_code": 0,
+  "result_exists": true,
+  "sha256": "942065870b3e493e86b3fd503a597f2236b4b7c49073c3c02d5529b3adf4f530"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c4.json

@@ -0,0 +1,381 @@
+{
+  "metadata": {
+    "version": "0.4.29",
+    "engine": "vllm",
+    "model": "glm53-flash-trellismx-p8-k45",
+    "server": "127.0.0.1:8001",
+    "timestamp": "2026-09-09T13:58:50.906147",
+    "decode_mode": "duration",
+    "primary_decode_layer": "sustained_decode",
+    "duration_per_test": 20.0,
+    "request_count": 0,
+    "warmup_request_count": 0,
+    "run_burst": false,
+    "prefill_mode": "skipped",
+    "standalone_prefill": false,
+    "prefill_only": false,
+    "skip_prefill": true,
+    "burst_e2e_status": "not_run_use_--run-burst",
+    "burst_request_count": 0,
+    "burst_warmup_request_count": 0,
+    "burst_requests_per_concurrency": 5,
+    "decode_warmup_seconds": 3.0,
+    "decode_warmup_context": 8192,
+    "decode_warmup_concurrency": 1,
+    "cell_warmup_timeout_seconds": 180.0,
+    "cell_warmup_timeout_policy": "<=32k:60s,64k:120s,>=128k:180s when override is 0",
+    "show_capacity_limited_values": false,
+    "max_tokens": 8192,
+    "temperature": 0.0,
+    "ignore_eos": true,
+    "max_total_tokens": 29540352,
+    "dcp_size": 0,
+    "metrics_available": true,
+    "metrics_warning": "",
+    "concurrency_levels": [
+      4
+    ],
+    "context_lengths": [
+      8192
+    ],
+    "startup_diagnostics_available": true,
+    "nvidia_p2p_override_effective": true,
+    "p2pmark_status": "not_run",
+    "amd_fabric_status": "not_run"
+  },
+  "startup_diagnostics": {
+    "version": "0.4.29",
+    "server_url": "http://127.0.0.1:8001",
+    "hostname": "<host>",
+    "uname": "Linux <host> 6.18.7-76061807-generic #202601231045~1769703228~24.04~cb87b5b SMP PREEMPT_DYNAMIC Thu J x86_64 x86_64 x86_64 GNU/Linux",
+    "env": {},
+    "args": {
+      "concurrency": "4",
+      "contexts": "8k",
+      "max_tokens": 8192,
+      "duration": 20.0,
+      "request_count": 0,
+      "run_burst": false,
+      "standalone_prefill": false,
+      "prefill_only": false,
+      "skip_prefill": true,
+      "prefill_contexts": "8k,64k,128k",
+      "prefill_metric": "client",
+      "dcp_size": 0,
+      "kv_budget": 0
+    },
+    "nvidia_p2p_override": {
+      "effective": true,
+      "configured": true,
+      "params_path": "/proc/driver/nvidia/params",
+      "params_available": true,
+      "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+      "modprobe_available": true,
+      "runtime": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1",
+        "DmaRemapPeerMmio": "1"
+      },
+      "expected": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1"
+      },
+      "missing": [],
+      "mismatched": {},
+      "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+      "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+      "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+    },
+    "p2pmark": {
+      "status": "not_run"
+    },
+    "amd_fabric": {
+      "status": "not_run"
+    },
+    "nvidia_smi_query": {
+      "cmd": [
+        "nvidia-smi",
+        "--query-gpu=index,name,driver_version,pci.bus_id,pcie.link.gen.current,pcie.link.width.current,power.limit",
+        "--format=csv,noheader,nounits"
+      ],
+      "returncode": 0,
+      "stdout": "0, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:01:00.0, 1, 16, 300.00\n1, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:21:00.0, 1, 16, 300.00\n2, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:81:00.0, 1, 16, 300.00\n3, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:C1:00.0, 1, 16, 300.00",
+      "stderr": ""
+    },
+    "nvidia_smi_topo": {
+      "cmd": [
+        "nvidia-smi",
+        "topo",
+        "-m"
+      ],
+      "returncode": 0,
+      "stdout": "\u001b[4mGPU0\tGPU1\tGPU2\tGPU3\tCPU Affinity\tNUMA Affinity\tGPU NUMA ID\u001b[0m\nGPU0\t X \tNODE\tNODE\tNODE\t0-47\t0\t\tN/A\nGPU1\tNODE\t X \tNODE\tNODE\t0-47\t0\t\tN/A\nGPU2\tNODE\tNODE\t X \tNODE\t0-47\t0\t\tN/A\nGPU3\tNODE\tNODE\tNODE\t X \t0-47\t0\t\tN/A\n\nLegend:\n\n  X    = Self\n  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)\n  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node\n  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)\n  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)\n  PIX  = Connection traversing at most a single PCIe bridge\n  NV#  = Connection traversing a bonded set of # NVLinks",
+      "stderr": ""
+    }
+  },
+  "nvidia_p2p_override": {
+    "effective": true,
+    "configured": true,
+    "params_path": "/proc/driver/nvidia/params",
+    "params_available": true,
+    "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+    "modprobe_available": true,
+    "runtime": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1",
+      "DmaRemapPeerMmio": "1"
+    },
+    "expected": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1"
+    },
+    "missing": [],
+    "mismatched": {},
+    "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+    "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+    "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+  },
+  "p2pmark": {
+    "status": "not_run"
+  },
+  "amd_fabric": {
+    "status": "not_run"
+  },
+  "hardware_run_summary": {
+    "samples": 18,
+    "duration_seconds": 40.845,
+    "gpu_count": 4,
+    "cpu_util_avg_pct": 9.89,
+    "cpu_temp_max_c": 77.25,
+    "gpu_util_avg_pct": 84.51,
+    "gpu_util_max_pct": 100.0,
+    "mem_util_avg_pct": 38.51,
+    "mem_util_max_pct": 63.0,
+    "temp_avg_c": 51.68,
+    "temp_max_c": 71.0,
+    "power_total_avg_w": 1042.47,
+    "power_total_max_w": 1177.98,
+    "power_limit_total_w": 1200.0,
+    "vram_used_avg_mb": 384450.0,
+    "vram_used_max_mb": 384450.0,
+    "vram_total_mb": 391548.0,
+    "vram_used_avg_pct": 98.19,
+    "vram_used_max_pct": 98.19,
+    "pcie_rx_avg_mb_s": 9780.89,
+    "pcie_rx_max_mb_s": 17065.0,
+    "pcie_tx_avg_mb_s": 9922.72,
+    "pcie_tx_max_mb_s": 21482.0
+  },
+  "event_log": [
+    "13:58:07 benchmark start engine=vllm",
+    "13:58:07 startup server=http://127.0.0.1:8001 model=glm53-flash-trellismx-p8-k45",
+    "13:58:07 startup decode concurrency=4 contexts=8k",
+    "13:58:07 startup NVIDIA P2P override: enabled: runtime NVIDIA P2P override matches expected RegistryDwords",
+    "13:58:07 startup engine vLLM 0.26.1rc0+glm53.flash.nvfp4.luke.clean.r1.vllme75bcfd.b12x58a046f models=['glm53-flash-trellismx-p8-k45']",
+    "13:58:07 startup KV cache budget from vLLM metrics: 29,540,352 tokens (3606 blocks x 2048; local 7,385,088 \u00d7 CP 4; CP source: local process)",
+    "13:58:07 startup model context length: 1,000,000 tokens",
+    "13:58:07 startup prefill tests: skipped",
+    "13:58:07 startup calibrating padding text run=tmxrepeatabc up_to=8k",
+    "13:58:07 startup context 8k: 50,558 chars (8,192 prompt tokens via /tokenize)",
+    "13:58:07 startup token targeting: /tokenize exact",
+    "13:58:07 startup startup preparation done",
+    "13:58:07 hardware monitor interval=2s",
+    "13:58:07 decode warmup start",
+    "13:58:07 decode warmup start C=1 ctx=8k 3s",
+    "13:58:07 cell start C=1 ctx=8k",
+    "13:58:12 ready C=1 ctx=8k running_reqs=1/1, queue_reqs=0, active_streams=1/1, stable=3.0s",
+    "13:58:15 cell done C=1 ctx=8k 192.4 tok/s",
+    "13:58:15 decode warmup done C=1 ctx=8k",
+    "13:58:17 cell start C=4 ctx=8k",
+    "13:58:28 ready C=4 ctx=8k running_reqs=4/4, queue_reqs=0, active_streams=4/4, stable=3.0s",
+    "13:58:48 cell done C=4 ctx=8k 318.6 tok/s"
+  ],
+  "prefill": {},
+  "results": [
+    {
+      "concurrency": 4,
+      "context_tokens": 8192,
+      "benchmark_mode": "duration",
+      "request_count_target": 0,
+      "warmup_request_count": 0,
+      "measurement_seconds": 19.989926,
+      "measurement_wall_seconds": 20.001037,
+      "client_output_tokens": 6368,
+      "server_output_tokens": 6368,
+      "aggregate_source": "openai_continuous_usage",
+      "aggregate_tps": 318.5604660251159,
+      "per_request_avg_tps": 79.64011650627897,
+      "ttft_avg": 2.812114993983414,
+      "ttft_p50": 1.9417693634750322,
+      "ttft_p90": 5.361200464423747,
+      "ttft_p99": 6.6800836534705,
+      "time_to_second_token_avg": 0.03318154823500663,
+      "time_to_second_token_p50": 0.039073001011274755,
+      "time_to_second_token_p90": 0.039100471721030775,
+      "time_to_second_token_p99": 0.039105877357069406,
+      "request_latency_avg": 0.0,
+      "request_latency_p50": 0.0,
+      "request_latency_p90": 0.0,
+      "request_latency_p99": 0.0,
+      "inter_token_latency_avg": 0.012973218425391987,
+      "inter_token_latency_p50": 0.012931780002378697,
+      "inter_token_latency_p90": 0.013394779352258331,
+      "inter_token_latency_p99": 0.013571113701463767,
+      "output_tps_per_user_avg": 77.15808171529739,
+      "output_tps_per_user_p50": 77.3288908019984,
+      "output_tps_per_user_p90": 79.48553383788095,
+      "output_tps_per_user_p99": 80.30392075134469,
+      "e2e_output_tps_per_user_avg": 0.0,
+      "e2e_output_tps_per_user_p50": 0.0,
+      "e2e_output_tps_per_user_p90": 0.0,
+      "e2e_output_tps_per_user_p99": 0.0,
+      "chunk_inter_token_latency_avg": 0.03485549465867402,
+      "chunk_inter_token_latency_p50": 0.03539645546581502,
+      "chunk_inter_token_latency_p90": 0.035495686222436366,
+      "chunk_inter_token_latency_p99": 0.03552552531765144,
+      "input_seq_len_avg": 8192.0,
+      "output_seq_len_avg": 2136.75,
+      "output_seq_len_p50": 2211.5,
+      "output_seq_len_p90": 2214.4,
+      "output_seq_len_p99": 2214.94,
+      "request_count": 4,
+      "completed_request_count": 0,
+      "request_samples": [
+        {
+          "ttft": 0.5382950189523399,
+          "time_to_second_token": 0.015473712934181094,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.013590706406931037,
+          "chunk_inter_token_latency_avg": 0.03552884077267534,
+          "input_tokens": 8192,
+          "output_tokens": 2210,
+          "output_tps_per_user": 73.57969262657433,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        },
+        {
+          "ttft": 1.9418736780062318,
+          "time_to_second_token": 0.039059544913470745,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.012937616224688685,
+          "chunk_inter_token_latency_avg": 0.035418325605212096,
+          "input_tokens": 8192,
+          "output_tokens": 2213,
+          "output_tps_per_user": 77.29399161583669,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        },
+        {
+          "ttft": 1.9416650489438325,
+          "time_to_second_token": 0.039086457109078765,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.012925943780068708,
+          "chunk_inter_token_latency_avg": 0.03537458532641795,
+          "input_tokens": 8192,
+          "output_tokens": 2215,
+          "output_tps_per_user": 77.36378998816011,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        },
+        {
+          "ttft": 6.826626230031252,
+          "time_to_second_token": 0.03910647798329592,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.012438607289879515,
+          "chunk_inter_token_latency_avg": 0.03310022693039068,
+          "input_tokens": 8192,
+          "output_tokens": 1909,
+          "output_tps_per_user": 80.39485263061844,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        }
+      ],
+      "total_tokens": 6368,
+      "wall_time": 31.101988177979365,
+      "num_completed": 4,
+      "num_errors": 0,
+      "server_gen_throughput": 318.31328873956187,
+      "server_utilization": 0.024410540915395296,
+      "server_spec_accept_rate": 0.5645161290322581,
+      "server_spec_accept_length": 0.0,
+      "avg_running_reqs": 4,
+      "max_running_reqs": 4,
+      "effective_concurrency": 4,
+      "avg_queue_reqs": 0,
+      "max_queue_reqs": 0,
+      "queue_fraction": 0.0,
+      "underfilled": false,
+      "warmup_timed_out": false,
+      "warmup_duration": 11.078,
+      "ready_reason": "running_reqs=4/4, queue_reqs=0, active_streams=4/4, stable=3.0s",
+      "timeout_reason": "",
+      "capacity_limited": false,
+      "hardware_summary": {
+        "samples": 9,
+        "duration_seconds": 19.215,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.82,
+        "cpu_temp_max_c": 77.25,
+        "gpu_util_avg_pct": 100.0,
+        "gpu_util_max_pct": 100.0,
+        "mem_util_avg_pct": 43.14,
+        "mem_util_max_pct": 51.0,
+        "temp_avg_c": 56.08,
+        "temp_max_c": 71.0,
+        "power_total_avg_w": 1174.33,
+        "power_total_max_w": 1177.98,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384450.0,
+        "vram_used_max_mb": 384450.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.19,
+        "vram_used_max_pct": 98.19,
+        "pcie_rx_avg_mb_s": 11085.78,
+        "pcie_rx_max_mb_s": 11507.0,
+        "pcie_tx_avg_mb_s": 11123.33,
+        "pcie_tx_max_mb_s": 11498.0
+      }
+    }
+  ],
+  "summary_table": {
+    "8192": {
+      "4": 318.5604660251159
+    }
+  },
+  "burst_results": [],
+  "burst_summary_table": {},
+  "methodology": {
+    "prefill": {
+      "name": "Prefill",
+      "present": false,
+      "mode": "skipped",
+      "formula": "prompt_tokens / TTFT",
+      "notes": "Default mode records the required decode scout request for each non-zero decode context, so normal runs do not pay for a separate prefill phase. Standalone mode repeats cold-prefill samples. Prometheus prefill counters, when available and uncontaminated, are stored as validation."
+    },
+    "sustained_decode": {
+      "name": "Sustained Decode",
+      "present": true,
+      "formula": "OpenAI stream usage completion_tokens per measured window; client chunk fallback only when continuous usage is unavailable",
+      "notes": "Duration-based steady-state cell after warmup. This is the main tuning/regression signal for kernels, NCCL, DCP, MTP, and scheduling. Prometheus metrics are stored as validation and scheduler state, not the default headline."
+    },
+    "burst_e2e_decode": {
+      "name": "Burst / E2E Decode",
+      "present": false,
+      "status": "not run; use --run-burst",
+      "formula": "sum(completion_tokens) / profiling_wall_time",
+      "notes": "Finite client-facing request burst using OpenAI stream usage. It includes request admission, scheduling, prefill/cache behavior, and completion."
+    }
+  }
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c4.log

@@ -0,0 +1,109 @@
+{"diagnostic_wrapper": "fixed-prefix-greedy", "benchmark_sha256": "7239958032ec10781d7db3efca23a7602bd9aeb94455a723e2634ab7d6fe546a", "run_id": "tmxrepeatabc", "temperature": 0.0, "scope": "Separate diagnostic, not model-default throughput"}
+
+New version available: v0.6.2 (current: v0.4.29)
+Upgrade and restart? [Y/n]: Skipping update.
+
+╭──────────────────────────── NVIDIA P2P Override ─────────────────────────────╮
+│ Effective: yes                                                               │
+│ Configured file: yes (/etc/modprobe.d/nvidia-p2p-override.conf)              │
+│ Runtime: ForceP2P=0x11; RMForceP2PType=1; RMPcieP2PType=2;                   │
+│ GrdmaPciTopoCheckOverride=1; EnableResizableBar=1; DmaRemapPeerMmio=1        │
+╰──────────────────────────────────────────────────────────────────────────────╯
+╭─────────────────────────────── Configuration ────────────────────────────────╮
+│ LLM Inference Benchmark                                                      │
+│ Model: glm53-flash-trellismx-p8-k45 @ 127.0.0.1:8001                         │
+│ Decode concurrency: [4]                                                      │
+│ Decode contexts: ['8k']                                                      │
+│ Duration: 20.0s per decode test | Max tokens: 8192                           │
+│ Pre-decode warmup: C=1 max-runnable context for 3s                           │
+│ Prefill: skipped | Sustained decode: 1 cells                                 │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Engine: vLLM 0.26.1rc0+glm53.flash.nvfp4.luke.clean.r1.vllme75bcfd.b12x58a046f  
+Models: ['glm53-flash-trellismx-p8-k45']
+KV cache budget (vLLM metrics): 29,540,352 tokens (3606 blocks × 2048; local 
+7,385,088 × CP 4; CP source: local process)
+Model context length: 1,000,000 tokens
+Prefill tests: skipped
+Calibrating padding text (run=tmxrepeatabc, up to 8k)...
+  8k: 50,558 chars (8,192 prompt tokens via /tokenize)
+  Token targeting: /tokenize exact
+Done.
+
+
+
+llm-decode-bench v0.4.29
+╭────────────────────────────────── Phase 2 ───────────────────────────────────╮
+│ Sustained Decode                                                             │
+│ Steady-state decode throughput after the engine has admitted the requested   │
+│ concurrency and passed warmup. Use this as the main tuning/regression signal │
+│ for kernels, NCCL, DCP, MTP, and scheduler changes.                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Aggregate tok/s  + TTFT/ITL 
+╭────────────┬─────────────╮
+│ ctx \ conc │           4 │
+├────────────┼─────────────┤
+│ 8k         │ 318.6 2k/13 │
+╰────────────┴─────────────╯
+Sustained Decode: aggregate tok/s uses OpenAI stream usage by default 
+(continuous completion_tokens when the server supports it). Prometheus is kept 
+as validation/scheduler data.
+Aggregate source(s): openai_continuous_usage
+Per-Request tok/s    
+╭────────────┬──────╮
+│ ctx \ conc │    4 │
+├────────────┼──────┤
+│ 8k         │ 79.6 │
+╰────────────┴──────╯
+Client request      
+latency: p50 / p90  
+ms                  
+╭────────────┬─────╮
+│ ctx \ conc │   4 │
+├────────────┼─────┤
+│ 8k         │ —/— │
+╰────────────┴─────╯
+Aggregate cells show dim detail as TTFT ms / ITL ms for the same ctx/conc 
+coordinate. ITL is computed from observed generated tokens, including streams 
+stopped at the measurement boundary; a missing ITL means no stream produced at 
+least two measured output tokens. Per-request tok/s and request latency are 
+shown in separate per-cell matrices. Completion/sample counts and full 
+request-level distributions remain in JSON under request_samples.
+Sustained mode: client latency metrics explain request UX variance; aggregate 
+tok/s remains the primary throughput signal. 
+ITL=(last_token_time-first_token_time)/(output_tokens-1), user tok/s=1/ITL.
+Hardware Summary                                                                
+╭───┬─┬───────┬───────────┬───────┬─────────┬─────┬──────┬─────┬───────────────╮
+│ … │ │ mode  │ GPU avg/… │ Mem … │ W avg/… │ T … │ CPU… │ VR… │ PCIe rx/tx a… │
+├───┼─┼───────┼───────────┼───────┼─────────┼─────┼──────┼─────┼───────────────┤
+│ … │ │ sust… │  100/100% │   43% │ 1174/1… │ 71C │  77C │ 98… │   11086/11123 │
+╰───┴─┴───────┴───────────┴───────┴─────────┴─────┴──────┴─────┴───────────────╯
+╭─────────────────────── Whole-run GPU Power ───────────────────────╮
+│ avg 1,042 W | max 1,178 W | limit 1,200 W | over 41s | 18 samples │
+╰───────────────────────────────────────────────────────────────────╯
+Hardware summary is sampled from nvidia-smi during the measured part of each 
+cell. Whole-run GPU power is the sampled sum of GPU power draw across the 
+complete benchmark run, not wall-outlet system power. PCIe rx/tx is MB/s and is 
+a coarse live diagnostic, not a per-kernel NCCL profiler.
+
+╭────────────────────────────────── Phase 3 ───────────────────────────────────╮
+│ Burst / E2E Decode                                                           │
+│ Not run. Re-run with --run-burst to append a finite client-facing request    │
+│ burst after Sustained Decode. This is intentionally disabled by default      │
+│ because it adds another full decode matrix.                                  │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭────────────────────────────── Primary Summary ───────────────────────────────╮
+│ Primary matrices repeated last so the important numbers are visible without  │
+│ scrolling back through diagnostics.                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Aggregate decode tok/s
+╭────────────┬───────╮
+│ ctx \ conc │     4 │
+├────────────┼───────┤
+│ 8k         │ 318.6 │
+╰────────────┴───────╯
+
+Results saved to 
+<campaign>/tile-policy-route-
+hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode-c4.json
+{"fixed_run_id_calls": 1, "benchmark_unchanged": true}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode16k-c1-command.json

@@ -0,0 +1,29 @@
+[
+  "/usr/bin/python3",
+  "<campaign>/fixed-workload-candidate-comparison-01/bench_fixed.py",
+  "--host",
+  "127.0.0.1",
+  "--port",
+  "8001",
+  "--model",
+  "glm53-flash-trellismx-p8-k45",
+  "--duration",
+  "20",
+  "--max-tokens",
+  "8192",
+  "--token-targeting",
+  "exact",
+  "--display-mode",
+  "plain",
+  "--output",
+  "<campaign>/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode16k-c1.json",
+  "--contexts",
+  "16k",
+  "--concurrency",
+  "1",
+  "--skip-prefill",
+  "--cell-warmup-timeout-seconds",
+  "180",
+  "--temperature",
+  "0"
+]

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode16k-c1-cooldown.json

@@ -0,0 +1,84 @@
+{
+  "status": "passed",
+  "minimum_seconds": 90,
+  "temperature_ceiling_c": 55,
+  "stable_seconds": 30,
+  "samples": [
+    {
+      "elapsed_seconds": 4.409812390804291e-07,
+      "hardware_csv": "0, 65, 2317, 114.23, 300.00\n1, 45, 2857, 142.19, 300.00\n2, 55, 2317, 87.31, 300.00\n3, 47, 2850, 115.10, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 5.10895173298195,
+      "hardware_csv": "0, 62, 2272, 84.00, 300.00\n1, 43, 2610, 87.94, 300.00\n2, 53, 2272, 76.78, 300.00\n3, 46, 2610, 99.63, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 10.204709953861311,
+      "hardware_csv": "0, 61, 2272, 80.40, 300.00\n1, 42, 2610, 87.23, 300.00\n2, 51, 2272, 75.83, 300.00\n3, 46, 2610, 99.02, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 15.31385334697552,
+      "hardware_csv": "0, 60, 2272, 78.87, 300.00\n1, 41, 1567, 79.14, 300.00\n2, 50, 2272, 71.85, 300.00\n3, 45, 2610, 99.15, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 20.429006435908377,
+      "hardware_csv": "0, 57, 180, 15.58, 300.00\n1, 40, 180, 18.69, 300.00\n2, 47, 180, 11.06, 300.00\n3, 43, 180, 32.80, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 25.571171382907778,
+      "hardware_csv": "0, 56, 180, 14.70, 300.00\n1, 39, 180, 15.84, 300.00\n2, 46, 180, 10.54, 300.00\n3, 43, 180, 23.63, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 30.708020013058558,
+      "hardware_csv": "0, 55, 172, 13.55, 300.00\n1, 39, 180, 15.70, 300.00\n2, 45, 180, 9.92, 300.00\n3, 42, 180, 23.10, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 35.83965059998445,
+      "hardware_csv": "0, 54, 180, 14.04, 300.00\n1, 38, 180, 15.98, 300.00\n2, 44, 180, 9.46, 300.00\n3, 41, 180, 22.49, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 40.98782790102996,
+      "hardware_csv": "0, 53, 180, 13.85, 300.00\n1, 38, 180, 15.67, 300.00\n2, 43, 180, 9.30, 300.00\n3, 41, 180, 22.45, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 46.131446681916714,
+      "hardware_csv": "0, 52, 180, 13.31, 300.00\n1, 37, 180, 15.28, 300.00\n2, 42, 180, 9.28, 300.00\n3, 40, 180, 22.12, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 51.267235012026504,
+      "hardware_csv": "0, 51, 180, 12.31, 300.00\n1, 37, 180, 15.44, 300.00\n2, 42, 180, 8.71, 300.00\n3, 40, 180, 21.95, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 56.396320046857,
+      "hardware_csv": "0, 50, 180, 12.41, 300.00\n1, 36, 180, 15.80, 300.00\n2, 41, 180, 8.51, 300.00\n3, 39, 180, 21.75, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 61.53433372802101,
+      "hardware_csv": "0, 50, 180, 12.15, 300.00\n1, 36, 180, 15.55, 300.00\n2, 40, 180, 8.20, 300.00\n3, 39, 172, 21.31, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 66.66481172991917,
+      "hardware_csv": "0, 49, 180, 11.46, 300.00\n1, 36, 180, 15.34, 300.00\n2, 40, 180, 8.24, 300.00\n3, 39, 180, 20.84, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 71.80251286691055,
+      "hardware_csv": "0, 49, 180, 11.22, 300.00\n1, 35, 180, 15.53, 300.00\n2, 39, 180, 7.69, 300.00\n3, 38, 180, 21.36, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 76.93311183084734,
+      "hardware_csv": "0, 48, 180, 11.21, 300.00\n1, 35, 180, 15.20, 300.00\n2, 39, 180, 7.64, 300.00\n3, 38, 180, 20.41, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 82.06933818990365,
+      "hardware_csv": "0, 48, 180, 11.10, 300.00\n1, 35, 180, 15.40, 300.00\n2, 38, 180, 7.73, 300.00\n3, 37, 180, 19.88, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 87.19939601491205,
+      "hardware_csv": "0, 47, 180, 11.27, 300.00\n1, 34, 180, 15.46, 300.00\n2, 38, 180, 6.99, 300.00\n3, 37, 180, 20.13, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 92.33550188690424,
+      "hardware_csv": "0, 47, 180, 11.04, 300.00\n1, 34, 180, 15.25, 300.00\n2, 37, 180, 7.22, 300.00\n3, 37, 180, 19.48, 300.00\n"
+    }
+  ]
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode16k-c1-fixed-workload.json

@@ -0,0 +1,5 @@
+{
+  "prefix": "tmxrepeatabc",
+  "temperature": 0,
+  "scope": "greedy fixed-prompt diagnostic"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode16k-c1-receipt.json

@@ -0,0 +1,5 @@
+{
+  "exit_code": 0,
+  "result_exists": true,
+  "sha256": "4bbf9c5a3d25601e2e466d64f22106a5166dadc8cb31fff0dc5661af9ff55573"
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode16k-c1.json

@@ -0,0 +1,345 @@
+{
+  "metadata": {
+    "version": "0.4.29",
+    "engine": "vllm",
+    "model": "glm53-flash-trellismx-p8-k45",
+    "server": "127.0.0.1:8001",
+    "timestamp": "2026-09-09T14:01:02.744264",
+    "decode_mode": "duration",
+    "primary_decode_layer": "sustained_decode",
+    "duration_per_test": 20.0,
+    "request_count": 0,
+    "warmup_request_count": 0,
+    "run_burst": false,
+    "prefill_mode": "skipped",
+    "standalone_prefill": false,
+    "prefill_only": false,
+    "skip_prefill": true,
+    "burst_e2e_status": "not_run_use_--run-burst",
+    "burst_request_count": 0,
+    "burst_warmup_request_count": 0,
+    "burst_requests_per_concurrency": 5,
+    "decode_warmup_seconds": 3.0,
+    "decode_warmup_context": 16384,
+    "decode_warmup_concurrency": 1,
+    "cell_warmup_timeout_seconds": 180.0,
+    "cell_warmup_timeout_policy": "<=32k:60s,64k:120s,>=128k:180s when override is 0",
+    "show_capacity_limited_values": false,
+    "max_tokens": 8192,
+    "temperature": 0.0,
+    "ignore_eos": true,
+    "max_total_tokens": 29540352,
+    "dcp_size": 0,
+    "metrics_available": true,
+    "metrics_warning": "",
+    "concurrency_levels": [
+      1
+    ],
+    "context_lengths": [
+      16384
+    ],
+    "startup_diagnostics_available": true,
+    "nvidia_p2p_override_effective": true,
+    "p2pmark_status": "not_run",
+    "amd_fabric_status": "not_run"
+  },
+  "startup_diagnostics": {
+    "version": "0.4.29",
+    "server_url": "http://127.0.0.1:8001",
+    "hostname": "<host>",
+    "uname": "Linux <host> 6.18.7-76061807-generic #202601231045~1769703228~24.04~cb87b5b SMP PREEMPT_DYNAMIC Thu J x86_64 x86_64 x86_64 GNU/Linux",
+    "env": {},
+    "args": {
+      "concurrency": "1",
+      "contexts": "16k",
+      "max_tokens": 8192,
+      "duration": 20.0,
+      "request_count": 0,
+      "run_burst": false,
+      "standalone_prefill": false,
+      "prefill_only": false,
+      "skip_prefill": true,
+      "prefill_contexts": "8k,64k,128k",
+      "prefill_metric": "client",
+      "dcp_size": 0,
+      "kv_budget": 0
+    },
+    "nvidia_p2p_override": {
+      "effective": true,
+      "configured": true,
+      "params_path": "/proc/driver/nvidia/params",
+      "params_available": true,
+      "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+      "modprobe_available": true,
+      "runtime": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1",
+        "DmaRemapPeerMmio": "1"
+      },
+      "expected": {
+        "ForceP2P": "0x11",
+        "RMForceP2PType": "1",
+        "RMPcieP2PType": "2",
+        "GrdmaPciTopoCheckOverride": "1",
+        "EnableResizableBar": "1"
+      },
+      "missing": [],
+      "mismatched": {},
+      "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+      "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+      "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+    },
+    "p2pmark": {
+      "status": "not_run"
+    },
+    "amd_fabric": {
+      "status": "not_run"
+    },
+    "nvidia_smi_query": {
+      "cmd": [
+        "nvidia-smi",
+        "--query-gpu=index,name,driver_version,pci.bus_id,pcie.link.gen.current,pcie.link.width.current,power.limit",
+        "--format=csv,noheader,nounits"
+      ],
+      "returncode": 0,
+      "stdout": "0, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:01:00.0, 1, 16, 300.00\n1, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:21:00.0, 1, 16, 300.00\n2, NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition, 610.57.04, 00000000:81:00.0, 1, 16, 300.00\n3, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 610.57.04, 00000000:C1:00.0, 1, 16, 300.00",
+      "stderr": ""
+    },
+    "nvidia_smi_topo": {
+      "cmd": [
+        "nvidia-smi",
+        "topo",
+        "-m"
+      ],
+      "returncode": 0,
+      "stdout": "\u001b[4mGPU0\tGPU1\tGPU2\tGPU3\tCPU Affinity\tNUMA Affinity\tGPU NUMA ID\u001b[0m\nGPU0\t X \tNODE\tNODE\tNODE\t0-47\t0\t\tN/A\nGPU1\tNODE\t X \tNODE\tNODE\t0-47\t0\t\tN/A\nGPU2\tNODE\tNODE\t X \tNODE\t0-47\t0\t\tN/A\nGPU3\tNODE\tNODE\tNODE\t X \t0-47\t0\t\tN/A\n\nLegend:\n\n  X    = Self\n  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)\n  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node\n  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)\n  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)\n  PIX  = Connection traversing at most a single PCIe bridge\n  NV#  = Connection traversing a bonded set of # NVLinks",
+      "stderr": ""
+    }
+  },
+  "nvidia_p2p_override": {
+    "effective": true,
+    "configured": true,
+    "params_path": "/proc/driver/nvidia/params",
+    "params_available": true,
+    "modprobe_path": "/etc/modprobe.d/nvidia-p2p-override.conf",
+    "modprobe_available": true,
+    "runtime": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1",
+      "DmaRemapPeerMmio": "1"
+    },
+    "expected": {
+      "ForceP2P": "0x11",
+      "RMForceP2PType": "1",
+      "RMPcieP2PType": "2",
+      "GrdmaPciTopoCheckOverride": "1",
+      "EnableResizableBar": "1"
+    },
+    "missing": [],
+    "mismatched": {},
+    "registry_dwords": "ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1",
+    "suggested_modprobe_line": "options nvidia NVreg_RegistryDwords=\"ForceP2P=0x11;RMForceP2PType=1;RMPcieP2PType=2;GrdmaPciTopoCheckOverride=1;EnableResizableBar=1\"",
+    "suggested_reload": "stop GPU workloads, then reload NVIDIA modules or reboot; the modprobe file alone is not enough until the nvidia module is reloaded"
+  },
+  "p2pmark": {
+    "status": "not_run"
+  },
+  "amd_fabric": {
+    "status": "not_run"
+  },
+  "hardware_run_summary": {
+    "samples": 16,
+    "duration_seconds": 36.095,
+    "gpu_count": 4,
+    "cpu_util_avg_pct": 9.84,
+    "cpu_temp_max_c": 76.38,
+    "gpu_util_avg_pct": 85.91,
+    "gpu_util_max_pct": 100.0,
+    "mem_util_avg_pct": 44.61,
+    "mem_util_max_pct": 61.0,
+    "temp_avg_c": 51.3,
+    "temp_max_c": 70.0,
+    "power_total_avg_w": 1046.13,
+    "power_total_max_w": 1167.75,
+    "power_limit_total_w": 1200.0,
+    "vram_used_avg_mb": 384450.0,
+    "vram_used_max_mb": 384450.0,
+    "vram_total_mb": 391548.0,
+    "vram_used_avg_pct": 98.19,
+    "vram_used_max_pct": 98.19,
+    "pcie_rx_avg_mb_s": 11433.5,
+    "pcie_rx_max_mb_s": 56886.0,
+    "pcie_tx_avg_mb_s": 11603.94,
+    "pcie_tx_max_mb_s": 47863.0
+  },
+  "event_log": [
+    "14:00:24 benchmark start engine=vllm",
+    "14:00:24 startup server=http://127.0.0.1:8001 model=glm53-flash-trellismx-p8-k45",
+    "14:00:24 startup decode concurrency=1 contexts=16k",
+    "14:00:24 startup NVIDIA P2P override: enabled: runtime NVIDIA P2P override matches expected RegistryDwords",
+    "14:00:24 startup engine vLLM 0.26.1rc0+glm53.flash.nvfp4.luke.clean.r1.vllme75bcfd.b12x58a046f models=['glm53-flash-trellismx-p8-k45']",
+    "14:00:24 startup KV cache budget from vLLM metrics: 29,540,352 tokens (3606 blocks x 2048; local 7,385,088 \u00d7 CP 4; CP source: local process)",
+    "14:00:24 startup model context length: 1,000,000 tokens",
+    "14:00:24 startup prefill tests: skipped",
+    "14:00:24 startup calibrating padding text run=tmxrepeatabc up_to=16k",
+    "14:00:24 startup context 16k: 102,112 chars (16,384 prompt tokens via /tokenize)",
+    "14:00:24 startup token targeting: /tokenize exact",
+    "14:00:24 startup startup preparation done",
+    "14:00:24 hardware monitor interval=2s",
+    "14:00:24 decode warmup start",
+    "14:00:24 decode warmup start C=1 ctx=16k 3s",
+    "14:00:24 cell start C=1 ctx=16k",
+    "14:00:30 ready C=1 ctx=16k running_reqs=1/1, queue_reqs=0, active_streams=1/1, stable=3.0s",
+    "14:00:33 cell done C=1 ctx=16k 197.6 tok/s",
+    "14:00:33 decode warmup done C=1 ctx=16k",
+    "14:00:35 cell start C=1 ctx=16k",
+    "14:00:40 ready C=1 ctx=16k running_reqs=1/1, queue_reqs=0, active_streams=1/1, stable=3.0s",
+    "14:01:00 cell done C=1 ctx=16k 216.6 tok/s"
+  ],
+  "prefill": {},
+  "results": [
+    {
+      "concurrency": 1,
+      "context_tokens": 16384,
+      "benchmark_mode": "duration",
+      "request_count_target": 0,
+      "warmup_request_count": 0,
+      "measurement_seconds": 19.992095,
+      "measurement_wall_seconds": 20.00022,
+      "client_output_tokens": 4331,
+      "server_output_tokens": 4331,
+      "aggregate_source": "openai_continuous_usage",
+      "aggregate_tps": 216.63562581500292,
+      "per_request_avg_tps": 216.63562581500292,
+      "ttft_avg": 0.506408799206838,
+      "ttft_p50": 0.506408799206838,
+      "ttft_p90": 0.506408799206838,
+      "ttft_p99": 0.506408799206838,
+      "time_to_second_token_avg": 0.01466080080717802,
+      "time_to_second_token_p50": 0.01466080080717802,
+      "time_to_second_token_p90": 0.01466080080717802,
+      "time_to_second_token_p99": 0.01466080080717802,
+      "request_latency_avg": 0.0,
+      "request_latency_p50": 0.0,
+      "request_latency_p90": 0.0,
+      "request_latency_p99": 0.0,
+      "inter_token_latency_avg": 0.00464640440922866,
+      "inter_token_latency_p50": 0.00464640440922866,
+      "inter_token_latency_p90": 0.00464640440922866,
+      "inter_token_latency_p99": 0.00464640440922866,
+      "output_tps_per_user_avg": 215.2201814404717,
+      "output_tps_per_user_p50": 215.2201814404717,
+      "output_tps_per_user_p90": 215.2201814404717,
+      "output_tps_per_user_p99": 215.2201814404717,
+      "e2e_output_tps_per_user_avg": 0.0,
+      "e2e_output_tps_per_user_p50": 0.0,
+      "e2e_output_tps_per_user_p90": 0.0,
+      "e2e_output_tps_per_user_p99": 0.0,
+      "chunk_inter_token_latency_avg": 0.013053681814469401,
+      "chunk_inter_token_latency_p50": 0.013053681814469401,
+      "chunk_inter_token_latency_p90": 0.013053681814469401,
+      "chunk_inter_token_latency_p99": 0.013053681814469401,
+      "input_seq_len_avg": 16384.0,
+      "output_seq_len_avg": 4954.0,
+      "output_seq_len_p50": 4954.0,
+      "output_seq_len_p90": 4954.0,
+      "output_seq_len_p99": 4954.0,
+      "request_count": 1,
+      "completed_request_count": 0,
+      "request_samples": [
+        {
+          "ttft": 0.506408799206838,
+          "time_to_second_token": 0.01466080080717802,
+          "latency": 0.0,
+          "inter_token_latency_avg": 0.00464640440922866,
+          "chunk_inter_token_latency_avg": 0.013053681814469401,
+          "input_tokens": 16384,
+          "output_tokens": 4954,
+          "output_tps_per_user": 215.2201814404717,
+          "e2e_output_tps_per_user": 0.0,
+          "completed": false
+        }
+      ],
+      "total_tokens": 4331,
+      "wall_time": 25.567179494071752,
+      "num_completed": 1,
+      "num_errors": 0,
+      "server_gen_throughput": 216.48234465686613,
+      "server_utilization": 0.00638002773925106,
+      "server_spec_accept_rate": 0.5757575757575758,
+      "server_spec_accept_length": 0.0,
+      "avg_running_reqs": 1,
+      "max_running_reqs": 1,
+      "effective_concurrency": 1,
+      "avg_queue_reqs": 0,
+      "max_queue_reqs": 0,
+      "queue_fraction": 0.0,
+      "underfilled": false,
+      "warmup_timed_out": false,
+      "warmup_duration": 5.56,
+      "ready_reason": "running_reqs=1/1, queue_reqs=0, active_streams=1/1, stable=3.0s",
+      "timeout_reason": "",
+      "capacity_limited": false,
+      "hardware_summary": {
+        "samples": 8,
+        "duration_seconds": 16.829,
+        "gpu_count": 4,
+        "cpu_util_avg_pct": 10.9,
+        "cpu_temp_max_c": 76.38,
+        "gpu_util_avg_pct": 99.0,
+        "gpu_util_max_pct": 99.0,
+        "mem_util_avg_pct": 54.84,
+        "mem_util_max_pct": 61.0,
+        "temp_avg_c": 54.31,
+        "temp_max_c": 69.0,
+        "power_total_avg_w": 1160.6,
+        "power_total_max_w": 1165.86,
+        "power_limit_total_w": 1200.0,
+        "vram_used_avg_mb": 384450.0,
+        "vram_used_max_mb": 384450.0,
+        "vram_total_mb": 391548.0,
+        "vram_used_avg_pct": 98.19,
+        "vram_used_max_pct": 98.19,
+        "pcie_rx_avg_mb_s": 9157.0,
+        "pcie_rx_max_mb_s": 9229.0,
+        "pcie_tx_avg_mb_s": 8978.0,
+        "pcie_tx_max_mb_s": 9037.0
+      }
+    }
+  ],
+  "summary_table": {
+    "16384": {
+      "1": 216.63562581500292
+    }
+  },
+  "burst_results": [],
+  "burst_summary_table": {},
+  "methodology": {
+    "prefill": {
+      "name": "Prefill",
+      "present": false,
+      "mode": "skipped",
+      "formula": "prompt_tokens / TTFT",
+      "notes": "Default mode records the required decode scout request for each non-zero decode context, so normal runs do not pay for a separate prefill phase. Standalone mode repeats cold-prefill samples. Prometheus prefill counters, when available and uncontaminated, are stored as validation."
+    },
+    "sustained_decode": {
+      "name": "Sustained Decode",
+      "present": true,
+      "formula": "OpenAI stream usage completion_tokens per measured window; client chunk fallback only when continuous usage is unavailable",
+      "notes": "Duration-based steady-state cell after warmup. This is the main tuning/regression signal for kernels, NCCL, DCP, MTP, and scheduling. Prometheus metrics are stored as validation and scheduler state, not the default headline."
+    },
+    "burst_e2e_decode": {
+      "name": "Burst / E2E Decode",
+      "present": false,
+      "status": "not run; use --run-burst",
+      "formula": "sum(completion_tokens) / profiling_wall_time",
+      "notes": "Finite client-facing request burst using OpenAI stream usage. It includes request admission, scheduling, prefill/cache behavior, and completion."
+    }
+  }
+}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode16k-c1.log

@@ -0,0 +1,109 @@
+{"diagnostic_wrapper": "fixed-prefix-greedy", "benchmark_sha256": "7239958032ec10781d7db3efca23a7602bd9aeb94455a723e2634ab7d6fe546a", "run_id": "tmxrepeatabc", "temperature": 0.0, "scope": "Separate diagnostic, not model-default throughput"}
+
+New version available: v0.6.2 (current: v0.4.29)
+Upgrade and restart? [Y/n]: Skipping update.
+
+╭──────────────────────────── NVIDIA P2P Override ─────────────────────────────╮
+│ Effective: yes                                                               │
+│ Configured file: yes (/etc/modprobe.d/nvidia-p2p-override.conf)              │
+│ Runtime: ForceP2P=0x11; RMForceP2PType=1; RMPcieP2PType=2;                   │
+│ GrdmaPciTopoCheckOverride=1; EnableResizableBar=1; DmaRemapPeerMmio=1        │
+╰──────────────────────────────────────────────────────────────────────────────╯
+╭─────────────────────────────── Configuration ────────────────────────────────╮
+│ LLM Inference Benchmark                                                      │
+│ Model: glm53-flash-trellismx-p8-k45 @ 127.0.0.1:8001                         │
+│ Decode concurrency: [1]                                                      │
+│ Decode contexts: ['16k']                                                     │
+│ Duration: 20.0s per decode test | Max tokens: 8192                           │
+│ Pre-decode warmup: C=1 max-runnable context for 3s                           │
+│ Prefill: skipped | Sustained decode: 1 cells                                 │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Engine: vLLM 0.26.1rc0+glm53.flash.nvfp4.luke.clean.r1.vllme75bcfd.b12x58a046f  
+Models: ['glm53-flash-trellismx-p8-k45']
+KV cache budget (vLLM metrics): 29,540,352 tokens (3606 blocks × 2048; local 
+7,385,088 × CP 4; CP source: local process)
+Model context length: 1,000,000 tokens
+Prefill tests: skipped
+Calibrating padding text (run=tmxrepeatabc, up to 16k)...
+  16k: 102,112 chars (16,384 prompt tokens via /tokenize)
+  Token targeting: /tokenize exact
+Done.
+
+
+
+llm-decode-bench v0.4.29
+╭────────────────────────────────── Phase 2 ───────────────────────────────────╮
+│ Sustained Decode                                                             │
+│ Steady-state decode throughput after the engine has admitted the requested   │
+│ concurrency and passed warmup. Use this as the main tuning/regression signal │
+│ for kernels, NCCL, DCP, MTP, and scheduler changes.                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Aggregate tok/s  + TTFT/ITL 
+╭────────────┬─────────────╮
+│ ctx \ conc │           1 │
+├────────────┼─────────────┤
+│ 16k        │ 216.6 506/5 │
+╰────────────┴─────────────╯
+Sustained Decode: aggregate tok/s uses OpenAI stream usage by default 
+(continuous completion_tokens when the server supports it). Prometheus is kept 
+as validation/scheduler data.
+Aggregate source(s): openai_continuous_usage
+Per-Request tok/s     
+╭────────────┬───────╮
+│ ctx \ conc │     1 │
+├────────────┼───────┤
+│ 16k        │ 216.6 │
+╰────────────┴───────╯
+Client request      
+latency: p50 / p90  
+ms                  
+╭────────────┬─────╮
+│ ctx \ conc │   1 │
+├────────────┼─────┤
+│ 16k        │ —/— │
+╰────────────┴─────╯
+Aggregate cells show dim detail as TTFT ms / ITL ms for the same ctx/conc 
+coordinate. ITL is computed from observed generated tokens, including streams 
+stopped at the measurement boundary; a missing ITL means no stream produced at 
+least two measured output tokens. Per-request tok/s and request latency are 
+shown in separate per-cell matrices. Completion/sample counts and full 
+request-level distributions remain in JSON under request_samples.
+Sustained mode: client latency metrics explain request UX variance; aggregate 
+tok/s remains the primary throughput signal. 
+ITL=(last_token_time-first_token_time)/(output_tokens-1), user tok/s=1/ITL.
+Hardware Summary                                                                
+╭───┬─┬───────┬───────────┬───────┬─────────┬─────┬──────┬─────┬───────────────╮
+│ … │ │ mode  │ GPU avg/… │ Mem … │ W avg/… │ T … │ CPU… │ VR… │ PCIe rx/tx a… │
+├───┼─┼───────┼───────────┼───────┼─────────┼─────┼──────┼─────┼───────────────┤
+│ … │ │ sust… │    99/99% │   55% │ 1161/1… │ 69C │  76C │ 98… │     9157/8978 │
+╰───┴─┴───────┴───────────┴───────┴─────────┴─────┴──────┴─────┴───────────────╯
+╭─────────────────────── Whole-run GPU Power ───────────────────────╮
+│ avg 1,046 W | max 1,168 W | limit 1,200 W | over 36s | 16 samples │
+╰───────────────────────────────────────────────────────────────────╯
+Hardware summary is sampled from nvidia-smi during the measured part of each 
+cell. Whole-run GPU power is the sampled sum of GPU power draw across the 
+complete benchmark run, not wall-outlet system power. PCIe rx/tx is MB/s and is 
+a coarse live diagnostic, not a per-kernel NCCL profiler.
+
+╭────────────────────────────────── Phase 3 ───────────────────────────────────╮
+│ Burst / E2E Decode                                                           │
+│ Not run. Re-run with --run-burst to append a finite client-facing request    │
+│ burst after Sustained Decode. This is intentionally disabled by default      │
+│ because it adds another full decode matrix.                                  │
+╰──────────────────────────────────────────────────────────────────────────────╯
+
+╭────────────────────────────── Primary Summary ───────────────────────────────╮
+│ Primary matrices repeated last so the important numbers are visible without  │
+│ scrolling back through diagnostics.                                          │
+╰──────────────────────────────────────────────────────────────────────────────╯
+Aggregate decode tok/s
+╭────────────┬───────╮
+│ ctx \ conc │     1 │
+├────────────┼───────┤
+│ 16k        │ 216.6 │
+╰────────────┴───────╯
+
+Results saved to 
+<campaign>/tile-policy-route-
+hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode16k-c1.json
+{"fixed_run_id_calls": 1, "benchmark_unchanged": true}

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode16k-c4-command.json

@@ -0,0 +1,29 @@
+[
+  "/usr/bin/python3",
+  "<campaign>/fixed-workload-candidate-comparison-01/bench_fixed.py",
+  "--host",
+  "127.0.0.1",
+  "--port",
+  "8001",
+  "--model",
+  "glm53-flash-trellismx-p8-k45",
+  "--duration",
+  "20",
+  "--max-tokens",
+  "8192",
+  "--token-targeting",
+  "exact",
+  "--display-mode",
+  "plain",
+  "--output",
+  "<campaign>/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode16k-c4.json",
+  "--contexts",
+  "16k",
+  "--concurrency",
+  "4",
+  "--skip-prefill",
+  "--cell-warmup-timeout-seconds",
+  "180",
+  "--temperature",
+  "0"
+]

results/speed-20260909/evidence/tile-policy-route-hoist-speed-01/results-01/tile-policy-route-hoist/rep-1/decode16k-c4-cooldown.json

@@ -0,0 +1,84 @@
+{
+  "status": "passed",
+  "minimum_seconds": 90,
+  "temperature_ceiling_c": 55,
+  "stable_seconds": 30,
+  "samples": [
+    {
+      "elapsed_seconds": 2.7008354663848877e-07,
+      "hardware_csv": "0, 64, 2317, 87.12, 300.00\n1, 44, 2857, 141.41, 300.00\n2, 53, 2310, 78.07, 300.00\n3, 48, 2850, 122.51, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 5.098574468167499,
+      "hardware_csv": "0, 61, 2272, 82.70, 300.00\n1, 43, 2610, 87.37, 300.00\n2, 51, 2272, 76.09, 300.00\n3, 46, 2610, 98.97, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 10.200626248959452,
+      "hardware_csv": "0, 60, 2272, 80.49, 300.00\n1, 42, 2610, 87.01, 300.00\n2, 50, 2272, 75.44, 300.00\n3, 45, 2610, 98.58, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 15.30647600395605,
+      "hardware_csv": "0, 58, 375, 56.32, 300.00\n1, 42, 2610, 86.17, 300.00\n2, 48, 360, 38.13, 300.00\n3, 44, 450, 89.40, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 20.424240899970755,
+      "hardware_csv": "0, 56, 180, 15.68, 300.00\n1, 40, 180, 19.04, 300.00\n2, 47, 180, 10.85, 300.00\n3, 43, 180, 27.39, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 25.558568245032802,
+      "hardware_csv": "0, 55, 180, 14.81, 300.00\n1, 39, 180, 15.38, 300.00\n2, 45, 180, 10.24, 300.00\n3, 42, 180, 22.72, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 30.690099691040814,
+      "hardware_csv": "0, 54, 180, 14.82, 300.00\n1, 39, 180, 15.36, 300.00\n2, 44, 180, 9.64, 300.00\n3, 42, 180, 21.73, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 35.82131854095496,
+      "hardware_csv": "0, 53, 180, 13.93, 300.00\n1, 38, 180, 15.41, 300.00\n2, 44, 180, 9.61, 300.00\n3, 41, 180, 21.98, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 40.95601170603186,
+      "hardware_csv": "0, 52, 180, 13.94, 300.00\n1, 37, 180, 15.42, 300.00\n2, 43, 172, 8.99, 300.00\n3, 41, 180, 21.86, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 46.08785715303384,
+      "hardware_csv": "0, 51, 180, 13.15, 300.00\n1, 37, 180, 15.31, 300.00\n2, 42, 180, 8.72, 300.00\n3, 40, 180, 20.87, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 51.22077400609851,
+      "hardware_csv": "0, 51, 180, 12.61, 300.00\n1, 37, 180, 15.49, 300.00\n2, 41, 180, 8.95, 300.00\n3, 40, 180, 20.90, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 56.35183930606581,
+      "hardware_csv": "0, 50, 180, 12.52, 300.00\n1, 36, 180, 15.29, 300.00\n2, 41, 180, 8.79, 300.00\n3, 39, 180, 20.81, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 61.484954212093726,
+      "hardware_csv": "0, 49, 180, 12.30, 300.00\n1, 36, 180, 15.19, 300.00\n2, 40, 180, 8.64, 300.00\n3, 39, 180, 21.02, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 66.62038749805652,
+      "hardware_csv": "0, 49, 180, 12.03, 300.00\n1, 35, 180, 15.15, 300.00\n2, 39, 180, 8.48, 300.00\n3, 38, 180, 20.27, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 71.75501630315557,
+      "hardware_csv": "0, 48, 180, 11.53, 300.00\n1, 35, 180, 15.19, 300.00\n2, 39, 180, 8.49, 300.00\n3, 38, 180, 20.84, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 76.89354794216342,
+      "hardware_csv": "0, 48, 180, 11.72, 300.00\n1, 35, 180, 15.54, 300.00\n2, 38, 180, 8.27, 300.00\n3, 38, 180, 20.24, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 82.0311095060315,
+      "hardware_csv": "0, 47, 180, 11.05, 300.00\n1, 35, 180, 15.10, 300.00\n2, 38, 180, 8.17, 300.00\n3, 37, 180, 19.59, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 87.16535470308736,
+      "hardware_csv": "0, 47, 180, 10.81, 300.00\n1, 34, 172, 15.16, 300.00\n2, 37, 180, 8.10, 300.00\n3, 37, 180, 19.50, 300.00\n"
+    },
+    {
+      "elapsed_seconds": 92.29969097394496,
+      "hardware_csv": "0, 46, 180, 11.20, 300.00\n1, 34, 180, 15.20, 300.00\n2, 37, 180, 7.83, 300.00\n3, 37, 180, 18.80, 300.00\n"
+    }
+  ]
+}