Add files using upload-large-folder tool

experiments/kv-cache-fp8/capacity/paced/N0024/aiperf_stdout.log

@@ -0,0 +1,281 @@
+[0m│  8,316.29 │   0.00 │ 207,339.55 │ 115,440.94 │  36,499.00 │     96.03 │ 24,857.70 │
+│      Effective Decode Throughput Per User │     46.05 │   0.00 │  62,572.93 │     138.78 │     112.37 │     22.59 │    150.19 │
+│                         (tokens/sec/user) │           │        │            │            │            │           │           │
+│     Effective Prefill Throughput Per User │  6,943.03 │   0.00 │ 132,845.67 │ 104,735.17 │  28,669.97 │      0.00 │ 21,145.52 │
+│                         (tokens/sec/user) │           │        │            │            │            │           │           │
+│                 Tokens In Flight (tokens) │ 48,327.53 │   0.00 │ 315,948.39 │ 214,713.30 │ 137,438.40 │ 32,663.75 │ 56,792.35 │
+│         Effective Latency (CO-aware) (ms) │  5,093.02 │ 591.34 │  32,479.47 │  22,164.45 │  14,553.76 │  2,637.58 │  5,486.47 │
+└───────────────────────────────────────────┴───────────┴────────┴────────────┴────────────┴────────────┴───────────┴───────────┘
+                                              NVIDIA AIPerf | LLM Metrics: Active                                               
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                 Metric ┃       avg ┃      min ┃        max ┃        p99 ┃        p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│  Active Decode Throughput (tokens/sec) │    173.04 │     0.00 │ 167,077.17 │     608.66 │     334.77 │    115.18 │    540.33 │
+│ Active Prefill Throughput (tokens/sec) │ 66,496.69 │ 5,725.08 │ 207,266.85 │ 158,832.65 │ 111,319.85 │ 63,387.17 │ 33,504.23 │
+│      Active Decode Throughput Per User │     77.16 │     0.00 │  62,572.93 │     201.55 │     115.52 │     78.95 │    188.14 │
+│                      (tokens/sec/user) │           │          │            │            │            │           │           │
+│     Active Prefill Throughput Per User │ 56,214.18 │ 5,725.08 │ 132,845.67 │ 124,404.90 │ 100,532.07 │ 51,569.08 │ 29,162.85 │
+│                      (tokens/sec/user) │           │          │            │            │            │           │           │
+│   Active Total Throughput (tokens/sec) │ 13,386.74 │    37.04 │ 207,339.55 │ 123,062.39 │  63,445.82 │    179.05 │ 30,442.84 │
+└────────────────────────────────────────┴───────────┴──────────┴────────────┴────────────┴────────────┴───────────┴───────────┘
+                                                NVIDIA AIPerf | LLM Metrics: Usage                                                
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                  Metric ┃           avg ┃      min ┃       max ┃       p99 ┃       p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│            Usage Prompt Tokens (tokens) │     26,096.19 │ 8,094.00 │ 57,151.00 │ 53,923.55 │ 44,485.00 │ 23,735.50 │ 11,914.06 │
+│ Usage Prompt Cache Read Tokens (tokens) │     25,485.13 │ 7,744.00 │ 56,256.00 │ 53,367.04 │ 43,942.40 │ 23,168.00 │ 11,955.15 │
+│        Usage Completion Tokens (tokens) │        329.09 │    20.00 │  1,958.00 │  1,594.85 │  1,072.50 │    139.00 │    394.96 │
+│             Usage Total Tokens (tokens) │     26,425.29 │ 8,180.00 │ 58,714.00 │ 54,515.65 │ 44,919.50 │ 24,041.00 │ 12,043.69 │
+│      Total Usage Prompt Tokens (tokens) │ 22,599,304.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│    Total Usage Prompt Cache Read Tokens │ 21,866,240.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                (tokens) ��               │          │           │           │           │           │           │
+│       Overall Usage Prompt Cache Read % │         96.76 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                     (%) │               │          │           │           │           │           │           │
+│  Total Usage Completion Tokens (tokens) │    284,993.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│       Total Usage Total Tokens (tokens) │ 22,884,297.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+└─────────────────────────────────────────┴───────────────┴──────────┴───────────┴───────────┴───────────┴───────────┴───────────┘
+                                                    NVIDIA AIPerf | LLM Metrics                                                     
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                        Metric ┃       avg ┃      min ┃       max ┃       p99 ┃       p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│                      Time to First Token (ms) │    497.65 │   308.41 │  2,226.30 │  1,432.47 │    704.07 │    413.02 │    207.07 │
+│                     Time to Second Token (ms) │     26.78 │     0.00 │    585.31 │    318.73 │     41.39 │     20.61 │     47.82 │
+│               Time to First Output Token (ms) │  2,099.82 │   321.86 │ 20,081.61 │ 15,095.14 │  5,898.46 │    728.17 │  3,194.44 │
+│                          Request Latency (ms) │  5,079.11 │   589.93 │ 32,477.77 │ 22,162.90 │ 14,544.16 │  2,623.38 │  5,489.93 │
+│                      Inter Token Latency (ms) │     14.09 │     3.11 │     39.09 │     27.07 │     18.58 │     13.33 │      3.86 │
+│              Output Token Throughput Per User │     75.61 │    25.58 │    321.57 │    111.91 │     98.42 │     75.03 │     19.14 │
+│                             (tokens/sec/user) │           │          │           │           │           │           │           │
+│ E2E Output Token Throughput (tokens/sec/user) │     59.65 │    21.12 │     97.80 │     94.87 │     77.95 │     59.41 │     14.35 │
+│               Output Sequence Length (tokens) │    329.09 │    20.00 │  1,958.00 │  1,594.85 │  1,072.50 │    139.00 │    394.96 │
+│                Input Sequence Length (tokens) │ 26,096.19 │ 8,094.00 │ 57,151.00 │ 53,923.55 │ 44,485.00 │ 23,735.50 │ 11,914.06 │
+│          Output Token Throughput (tokens/sec) │    103.57 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│           Input Token Throughput (tokens/sec) │  8,213.03 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│             Request Throughput (requests/sec) │      0.31 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                      Request Count (requests) │    866.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+└───────────────────────────────────────────────┴───────────┴──────────┴───────────┴───────────┴───────────┴───────────┴───────────┘
+
+
+No GPU telemetry data collected during the benchmarking run.
+
+╭─────────────────────────────── Output Sequence Length Mismatch Warning ───────────────────────────────╮
+│  786 of 866 requests (90.8%) have output length differing from requested by more than the threshold.  │
+│  Threshold (tokens): min(requested x 5%, 50)                                                          │
+│  Average mismatch: -73.4%                                                                             │
+│                                                                                                       │
+│  Why: Server hit EOS token before reaching requested output length.                                   │
+│                                                                                                       │
+│  Fix Options:                                                                                         │
+│    - --extra-inputs ignore_eos:true - Generate until max_tokens (vLLM, TensorRT-LLM)                  │
+│    - --extra-inputs min_tokens:<N> - Set minimum output length (vLLM, TensorRT-LLM, SGLang)           │
+│    - --use-server-token-count - Use server-reported token counts if tokenizer mismatch suspected      │
+│                                                                                                       │
+│  Diagnostics:                                                                                         │
+│    - Review profile_export.jsonl -> osl_mismatch_diff_pct for per-request values                      │
+│    - Adjust: AIPERF_METRICS_OSL_MISMATCH_PCT_THRESHOLD=5                                              │
+│    - Adjust: AIPERF_METRICS_OSL_MISMATCH_MAX_TOKEN_THRESHOLD=50                                       │
+╰───────────────────────────────────────────────────────────────────────────────────────────────────────╯
+
+CLI Command: aiperf profile --model 'glm-5.2' --url 'http://127.0.0.1:8391' --api-key '<redacted>' --endpoint-type 'chat' --streaming --tokenizer 'builtin' --custom-dataset-type 
+'dag_jsonl' --input-file 'results/capacity/sweep-paced-jul20/N0024/paced.jsonl' --extra-inputs '{"cache_salt": "2aa386a91b47e21885111b7efc90adba"}' --fixed-schedule --request-count 
+1000000 --random-seed 0 --use-server-token-count --artifact-dir 'results/capacity/sweep-paced-jul20/N0024' --export-level 'raw' --export-http-trace --slice-duration 60 
+--server-metrics-formats 'json' 'csv' 'jsonl' --ui 'simple'
+Benchmark Duration: 2751.63 sec
+CSV Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0024/profile_export_aiperf.csv
+JSON Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0024/profile_export_aiperf.json
+Raw Records: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0024/profile_export_raw.jsonl
+Server Metrics CSV Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0024/server_metrics_export.csv
+Server Metrics JSON Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0024/server_metrics_export.json
+Timeslice CSV Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0024/profile_export_aiperf_timeslices.csv
+Timeslice JSON Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0024/profile_export_aiperf_timeslices.json
+Log File: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0024/logs/aiperf.log
+
+
+Profiling: 0/866 |          |   0% [00:00<?]
+Profiling: 0/866 |          |   0% [00:00<?]
+                                            
+
+Profiling: 9/866 |          |   1% [01:28<2:19:41]
+Profiling: 9/866 |          |   1% [01:28<2:19:41]
+Profiling: 19/866 |▏         |   2% [02:28<1:45:20]
+Profiling: 19/866 |▏         |   2% [02:28<1:45:20]
+Profiling: 28/866 |▎         |   3% [02:54<1:16:03]
+Profiling: 28/866 |▎         |   3% [02:54<1:16:03]
+Profiling: 37/866 |▍         |   4% [03:30<1:07:31]
+Profiling: 37/866 |▍         |   4% [03:30<1:07:31]
+Profiling: 46/866 |▌         |   5% [04:02<1:00:19]
+Profiling: 46/866 |▌         |   5% [04:02<1:00:19]
+Profiling: 55/866 |▋         |   6% [04:40<58:47]  
+Profiling: 55/866 |▋         |   6% [04:40<58:47]
+Profiling: 66/866 |▊         |   8% [05:04<47:17]
+Profiling: 66/866 |▊         |   8% [05:04<47:17]
+Profiling: 75/866 |▊         |   9% [05:28<43:15]
+Profiling: 75/866 |▊         |   9% [05:28<43:15]
+Profiling: 84/866 |▉         |  10% [05:40<35:06]
+Profiling: 84/866 |▉         |  10% [05:40<35:06]
+Profiling: 93/866 |█         |  11% [05:52<29:25]
+Profiling: 93/866 |█         |  11% [05:52<29:25]
+Profiling: 102/866 |█▏        |  12% [06:14<29:42]
+Profiling: 102/866 |█▏        |  12% [06:14<29:42]
+Profiling: 111/866 |█▎        |  13% [06:34<28:56]
+Profiling: 111/866 |█▎        |  13% [06:34<28:56]
+Profiling: 121/866 |█▍        |  14% [07:10<33:46]
+Profiling: 121/866 |█▍        |  14% [07:10<33:46]
+Profiling: 130/866 |█▌        |  15% [07:38<34:46]
+Profiling: 130/866 |█▌        |  15% [07:38<34:46]
+Profiling: 139/866 |█▌        |  16% [07:58<32:09]
+Profiling: 139/866 |█▌        |  16% [07:58<32:09]
+Profiling: 148/866 |█▋        |  17% [08:16<29:26]
+Profiling: 148/866 |█▋        |  17% [08:16<29:26]
+Profiling: 158/866 |█▊        |  18% [08:42<29:33]
+Profiling: 158/866 |█▊        |  18% [08:42<29:33]
+Profiling: 167/866 |█▉        |  19% [09:04<28:58]
+Profiling: 167/866 |█▉        |  19% [09:04<28:58]
+Profiling: 178/866 |██        |  21% [09:24<25:56]
+Profiling: 178/866 |██        |  21% [09:24<25:56]
+Profiling: 187/866 |██▏       |  22% [09:42<24:45]
+Profiling: 187/866 |██▏       |  22% [09:42<24:45]
+Profiling: 197/866 |██▎       |  23% [10:00<23:03]
+Profiling: 197/866 |██▎       |  23% [10:00<23:03]
+Profiling: 207/866 |██▍       |  24% [10:12<19:47]
+Profiling: 207/866 |██▍       |  24% [10:12<19:47]
+Profiling: 217/866 |██▌       |  25% [10:32<20:08]
+Profiling: 217/866 |██▌       |  25% [10:32<20:08]
+Profiling: 226/866 |██▌       |  26% [10:46<18:56]
+Profiling: 226/866 |██▌       |  26% [10:46<18:56]
+Profiling: 236/866 |██▋       |  27% [10:56<16:08]
+Profiling: 236/866 |██▋       |  27% [10:56<16:08]
+Profiling: 245/866 |██▊       |  28% [11:02<13:20]
+Profiling: 245/866 |██▊       |  28% [11:02<13:20]
+Profiling: 255/866 |██▉       |  29% [11:12<12:12]
+Profiling: 255/866 |██▉       |  29% [11:12<12:12]
+Profiling: 266/866 |███       |  31% [11:24<11:38]
+Profiling: 266/866 |███       |  31% [11:24<11:38]
+Profiling: 276/866 |███▏      |  32% [11:36<11:33]
+Profiling: 276/866 |███▏      |  32% [11:36<11:33]
+Profiling: 286/866 |███▎      |  33% [11:58<14:19]
+Profiling: 286/866 |███▎      |  33% [11:58<14:19]
+Profiling: 295/866 |███▍      |  34% [12:12<14:17]
+Profiling: 295/866 |███▍      |  34% [12:12<14:17]
+Profiling: 305/866 |███▌      |  35% [12:20<12:02]
+Profiling: 305/866 |███▌      |  35% [12:20<12:02]
+Profiling: 315/866 |███▋      |  36% [12:32<11:34]
+Profiling: 315/866 |███▋      |  36% [12:32<11:34]
+Profiling: 324/866 |███▋      |  37% [12:38<09:52]
+Profiling: 324/866 |███▋      |  37% [12:38<09:52]
+Profiling: 333/866 |███▊      |  38% [12:48<09:45]
+Profiling: 333/866 |███▊      |  38% [12:48<09:45]
+Profiling: 342/866 |███▉      |  39% [13:04<11:19]
+Profiling: 342/866 |███▉      |  39% [13:04<11:19]
+Profiling: 351/866 |████      |  41% [13:14<10:39]
+Profiling: 351/866 |████      |  41% [13:14<10:39]
+Profiling: 360/866 |████▏     |  42% [13:34<12:55]
+Profiling: 360/866 |████▏     |  42% [13:34<12:55]
+Profiling: 369/866 |████▎     |  43% [13:44<11:39]
+Profiling: 369/866 |████▎     |  43% [13:44<11:39]
+Profiling: 378/866 |████▎     |  44% [14:02<12:52]
+Profiling: 378/866 |████▎     |  44% [14:02<12:52]
+Profiling: 387/866 |████▍     |  45% [14:16<12:34]
+Profiling: 387/866 |████▍     |  45% [14:16<12:34]
+Profiling: 397/866 |████▌     |  46% [14:32<12:22]
+Profiling: 397/866 |████▌     |  46% [14:32<12:22]
+Profiling: 407/866 |████▋     |  47% [14:46<11:40]
+Profiling: 407/866 |████▋     |  47% [14:46<11:40]
+Profiling: 416/866 |████▊     |  48% [15:02<11:59]
+Profiling: 416/866 |████▊     |  48% [15:02<11:59]
+Profiling: 425/866 |████▉     |  49% [15:18<12:08]
+Profiling: 425/866 |████▉     |  49% [15:18<12:08]
+Profiling: 435/866 |█████     |  50% [15:28<10:22]
+Profiling: 435/866 |█████     |  50% [15:28<10:22]
+Profiling: 444/866 |█████▏    |  51% [15:36<09:02]
+Profiling: 444/866 |█████▏    |  51% [15:36<09:02]
+Profiling: 453/866 |█████▏    |  52% [15:48<08:56]
+Profiling: 453/866 |█████▏    |  52% [15:48<08:56]
+Profiling: 463/866 |█████▎    |  53% [15:58<08:05]
+Profiling: 463/866 |█████▎    |  53% [15:58<08:05]
+Profiling: 472/866 |█████▍    |  55% [16:04<06:53]
+Profiling: 472/866 |█████▍    |  55% [16:04<06:53]
+Profiling: 482/866 |█████▌    |  56% [16:14<06:36]
+Profiling: 482/866 |█████▌    |  56% [16:14<06:36]
+Profiling: 495/866 |█████▋    |  57% [16:28<06:29]
+Profiling: 495/866 |█████▋    |  57% [16:28<06:29]
+Profiling: 506/866 |█████▊    |  58% [16:42<06:42]
+Profiling: 506/866 |█████▊    |  58% [16:42<06:42]
+Profiling: 515/866 |█████▉    |  59% [16:52<06:31]
+Profiling: 515/866 |█████▉    |  59% [16:52<06:31]
+Profiling: 525/866 |██████    |  61% [17:04<06:29]
+Profiling: 525/866 |██████    |  61% [17:04<06:29]
+Profiling: 538/866 |██████▏   |  62% [17:14<05:31]
+Profiling: 538/866 |██████▏   |  62% [17:14<05:31]
+Profiling: 547/866 |██████▎   |  63% [17:24<05:30]
+Profiling: 547/866 |██████▎   |  63% [17:24<05:30]
+Profiling: 557/866 |██████▍   |  64% [17:42<06:29]
+Profiling: 557/866 |██████▍   |  64% [17:42<06:29]
+Profiling: 567/866 |██████▌   |  65% [18:00<07:04]
+Profiling: 567/866 |██████▌   |  65% [18:00<07:04]
+Profiling: 577/866 |██████▋   |  67% [18:10<06:14]
+Profiling: 577/866 |██████▋   |  67% [18:10<06:14]
+Profiling: 587/866 |██████▊   |  68% [18:20<05:36]
+Profiling: 587/866 |██████▊   |  68% [18:20<05:36]
+Profiling: 597/866 |██████▉   |  69% [18:30<05:08]
+Profiling: 597/866 |██████▉   |  69% [18:30<05:08]
+Profiling: 607/866 |███████   |  70% [18:48<05:47]
+Profiling: 607/866 |███████   |  70% [18:48<05:47]
+Profiling: 616/866 |███████   |  71% [19:08<06:36]
+Profiling: 616/866 |███████   |  71% [19:08<06:36]
+Profiling: 626/866 |███████▏  |  72% [19:26<06:36]
+Profiling: 626/866 |███████▏  |  72% [19:26<06:36]
+Profiling: 635/866 |███████▎  |  73% [19:34<05:31]
+Profiling: 635/866 |███████▎  |  73% [19:34<05:31]
+Profiling: 645/866 |███████▍  |  74% [19:48<05:14]
+Profiling: 645/866 |███████▍  |  74% [19:48<05:14]
+Profiling: 655/866 |███████▌  |  76% [20:04<05:12]
+Profiling: 655/866 |███████▌  |  76% [20:04<05:12]
+Profiling: 666/866 |███████▋  |  77% [20:24<05:17]
+Profiling: 666/866 |███████▋  |  77% [20:24<05:17]
+Profiling: 675/866 |███████▊  |  78% [20:44<05:36]
+Profiling: 675/866 |███████▊  |  78% [20:44<05:36]
+Profiling: 684/866 |███████▉  |  79% [21:04<05:44]
+Profiling: 684/866 |███████▉  |  79% [21:04<05:44]
+Profiling: 693/866 |████████  |  80% [21:20<05:21]
+Profiling: 693/866 |████████  |  80% [21:20<05:21]
+Profiling: 704/866 |████████▏ |  81% [21:44<05:18]
+Profiling: 704/866 |████████▏ |  81% [21:44<05:18]
+Profiling: 713/866 |████████▏ |  82% [22:06<05:21]
+Profiling: 713/866 |████████▏ |  82% [22:06<05:21]
+Profiling: 722/866 |████████▎ |  83% [22:32<05:35]
+Profiling: 722/866 |████████▎ |  83% [22:32<05:35]
+Profiling: 731/866 |████████▍ |  84% [23:00<05:44]
+Profiling: 731/866 |████████▍ |  84% [23:00<05:44]
+Profiling: 740/866 |████████▌ |  85% [23:18<05:01]
+Profiling: 740/866 |████████▌ |  85% [23:18<05:01]
+Profiling: 749/866 |████████▋ |  86% [23:34<04:18]
+Profiling: 749/866 |████████▋ |  86% [23:34<04:18]
+Profiling: 758/866 |████████▊ |  88% [24:18<05:24]
+Profiling: 758/866 |████████▊ |  88% [24:18<05:24]
+Profiling: 767/866 |████████▊ |  89% [25:50<08:30]
+Profiling: 767/866 |████████▊ |  89% [25:50<08:30]
+Profiling: 776/866 |████████▉ |  90% [26:12<06:31]
+Profiling: 776/866 |████████▉ |  90% [26:12<06:31]
+Profiling: 785/866 |█████████ |  91% [26:24<04:39]
+Profiling: 785/866 |█████████ |  91% [26:24<04:39]
+Profiling: 794/866 |█████████▏|  92% [26:48<03:51]
+Profiling: 794/866 |█████████▏|  92% [26:48<03:51]
+Profiling: 803/866 |█��███████▎|  93% [29:26<07:53]
+Profiling: 803/866 |█████████▎|  93% [29:26<07:53]
+Profiling: 812/866 |█████████▍|  94% [30:20<06:21]
+Profiling: 812/866 |█████████▍|  94% [30:20<06:21]
+Profiling: 821/866 |█████████▍|  95% [31:26<05:21]
+Profiling: 821/866 |█████████▍|  95% [31:26<05:21]
+Profiling: 830/866 |█████████▌|  96% [35:22<07:43]
+Profiling: 830/866 |█████████▌|  96% [35:22<07:43]
+Profiling: 839/866 |█████████▋|  97% [37:12<05:42]
+Profiling: 839/866 |█████████▋|  97% [37:12<05:42]
+Profiling: 848/866 |█████████▊|  98% [40:06<04:24]
+Profiling: 848/866 |█████████▊|  98% [40:06<04:24]
+Profiling: 857/866 |█████████▉|  99% [42:30<02:15]
+Profiling: 857/866 |█████████▉|  99% [42:30<02:15]
+Profiling: 866/866 |██████████| 100% [45:51<00:00]
+Profiling: 866/866 |██████████| 100% [45:51<00:00]
+Processing Records: 0/866 |          |   0% [00:00<?]
+Processing Records: 866/866 |██████████| 100% [00:00<00:00]
+Processing Records: 866/866 |██████████| 100% [00:00<00:00]

experiments/kv-cache-fp8/capacity/paced/N0024/manifest.json

@@ -0,0 +1,109 @@
+{
+  "concurrency": 24,
+  "mode": "paced",
+  "cache_salt": "2aa386a91b47e21885111b7efc90adba",
+  "command": [
+    "aiperf",
+    "profile",
+    "--model",
+    "glm-5.2",
+    "--url",
+    "http://127.0.0.1:8391",
+    "--api-key",
+    "<key>",
+    "--endpoint-type",
+    "chat",
+    "--streaming",
+    "--tokenizer",
+    "builtin",
+    "--custom-dataset-type",
+    "dag_jsonl",
+    "--input-file",
+    "results/capacity/sweep-paced-jul20/N0024/paced.jsonl",
+    "--extra-inputs",
+    "{\"cache_salt\": \"2aa386a91b47e21885111b7efc90adba\"}",
+    "--fixed-schedule",
+    "--request-count",
+    "1000000",
+    "--random-seed",
+    "0",
+    "--use-server-token-count",
+    "--artifact-dir",
+    "results/capacity/sweep-paced-jul20/N0024",
+    "--export-level",
+    "raw",
+    "--export-http-trace",
+    "--slice-duration",
+    "60",
+    "--server-metrics-formats",
+    "json",
+    "csv",
+    "jsonl",
+    "--ui",
+    "simple"
+  ],
+  "dataset_path": "data/capacity/corpus.jsonl",
+  "signature": {
+    "dataset_sha256": "0327d658a267715aebdab45b67b6e0d324616a0faeec8a75d09b8defcd0b2ac2",
+    "url": "http://127.0.0.1:8391",
+    "model": "glm-5.2",
+    "duration": 900,
+    "warmup": 120,
+    "grace": 300,
+    "seed": 0,
+    "request_ceiling": 1000000,
+    "python": "3.12.3",
+    "aiperf": {
+      "version": "0.11.0",
+      "rev": "73d0524f98f30199ca659a8ab586454a823c887d"
+    },
+    "server": {
+      "version": "0.5.15.post1",
+      "model_path": "/home/jl_fs/models/GLM-5.2-FP8",
+      "served_model_name": "glm-5.2",
+      "tp_size": 8,
+      "dp_size": 8,
+      "enable_cache_report": true,
+      "speculative_algorithm": "EAGLE",
+      "max_running_requests": 48,
+      "context_length": 262144,
+      "kv_cache_dtype": "fp8_e4m3",
+      "mem_fraction_static": 0.872,
+      "scheduler_pids": [
+        82149,
+        82150,
+        82151,
+        82152,
+        82153,
+        82154,
+        82155,
+        82156
+      ],
+      "max_total_num_tokens": 234176,
+      "max_model_len": 262144,
+      "model_id": "glm-5.2"
+    },
+    "evaluator_sha256": "3e48dfede43cbf47244f560547eda95d29c671d4fa23b203a5d220a14edf1c80",
+    "verifier_sha256": "6f84bee84368a72ea1445d0961185a0d5d2fd6ba8f8ec4c3fefac619b46a32aa",
+    "runner_sha256": "aa3e1a0f66bcfac6663b5d2e59a37cae9d7264dda6b4c13130a3edf403e77104",
+    "paced": {
+      "rhythm_sha256": "3dfc2e11cd63aa9b0a5ac61db594a9df320457a3a814e3e183c9e83d33490ffe",
+      "idle_cutoff": 300.0,
+      "pacer_sha256": "c9881fc93b09e8d11a3129e4639a889af1d913661a2c74d86ce59e762bd18558"
+    }
+  },
+  "started_utc": "2026-07-20T21:32:15Z",
+  "loadavg_start": [
+    1.51318359375,
+    0.81396484375,
+    0.6494140625
+  ],
+  "status": "ok",
+  "elapsed_s": 2762.3,
+  "loadavg_end": [
+    0.45068359375,
+    0.5791015625,
+    0.642578125
+  ],
+  "returncode": 0
+}

experiments/kv-cache-fp8/capacity/paced/N0024/paced.manifest.json

@@ -0,0 +1,45 @@
+{
+  "corpus": "data/capacity/corpus.jsonl",
+  "corpus_sha256": "0327d658a267715aebdab45b67b6e0d324616a0faeec8a75d09b8defcd0b2ac2",
+  "rhythm": "data/capacity/rhythm.json",
+  "rhythm_sha256": "3dfc2e11cd63aa9b0a5ac61db594a9df320457a3a814e3e183c9e83d33490ffe",
+  "rhythm_source_sha256": "29b6a19e751ff5230771519aab755f80a0f43a4ba9cf96b72d3a6a437ec99276",
+  "settings": {
+    "developers": 24,
+    "span_s": 900.0,
+    "seed": 0,
+    "idle_cutoff_s": 300.0
+  },
+  "arrival_model": "uniform over [0, span], seeded (traces publish no cross-session wall clock; open-loop by construction)",
+  "population_model": "one shuffled population per seed, N takes the prefix; per-session RNG streams keep a session's schedule/gaps identical across N",
+  "gap_model": "i.i.d. per turn from the conditional empirical distribution (marginal preserved; serial correlation not modeled)",
+  "idle_tail_dropped_frac": 0.041,
+  "python": "3.12.3",
+  "turns_total": 866,
+  "max_session_turns": 50,
+  "max_session_cap_tokens": 68222,
+  "last_arrival_s": 868.4,
+  "max_session_think_s": 1549.6,
+  "gap_percentiles_s": {
+    "source_conditional": {
+      "p10": 1.211,
+      "p25": 2.208,
+      "p50": 4.561,
+      "p75": 9.037,
+      "p90": 62.579,
+      "p95": 146.055,
+      "p99": 208.6
+    },
+    "sampled": {
+      "p10": 1.233,
+      "p25": 2.287,
+      "p50": 4.622,
+      "p75": 9.785,
+      "p90": 61.408,
+      "p95": 122.712,
+      "p99": 218.112
+    }
+  },
+  "ks_distance_sampled_vs_source": 0.0238,
+  "output_sha256": "7b728328d10be2bcad5ab5c58fdccd8404fdc9e365bbb426c75bd851b3a45d0c"
+}

experiments/kv-cache-fp8/capacity/paced/N0024/profile_export_aiperf.csv

@@ -0,0 +1,72 @@
+Metric,avg,min,max,sum,p1,p5,p10,p25,p50,p75,p90,p95,p99,std
+Active Decode Throughput (tokens/sec),173.04,0.00,167077.17,,,,,,115.18,,334.77,405.74,608.66,540.33
+Active Decode Throughput Per User (tokens/sec/user),77.16,0.00,62572.93,,,,,,78.95,,115.52,118.97,201.55,188.14
+Active Prefill Throughput (tokens/sec),66496.69,5725.08,207266.85,,,,,,63387.17,,111319.85,122990.85,158832.65,33504.23
+Active Prefill Throughput Per User (tokens/sec/user),56214.18,5725.08,132845.67,,,,,,51569.08,,100532.07,112473.50,124404.90,29162.85
+Active Total Throughput (tokens/sec),13386.74,37.04,207339.55,,,,,,179.05,,63445.82,88316.31,123062.39,30442.84
+Credit-to-Start Latency (ms),1.72,0.68,8.59,1489.43,1.00,1.11,1.18,1.29,1.47,1.72,2.42,3.62,5.66,0.87
+E2E Output Token Throughput (tokens/sec/user),59.65,21.12,97.80,51659.28,26.59,35.59,41.17,50.33,59.41,68.65,77.95,84.91,94.87,14.35
+Effective Concurrency,1.60,0.00,10.00,,,,,,1.00,,4.00,5.00,7.00,1.86
+Effective Decode Concurrency,1.45,0.00,9.00,,,,,,1.00,,4.00,5.00,7.00,1.73
+Effective Decode Throughput (tokens/sec),103.26,0.00,167077.17,,,,,,37.22,,273.93,347.14,517.71,425.94
+Effective Decode Throughput Per User (tokens/sec/user),46.05,0.00,62572.93,,,,,,22.59,,112.37,116.40,138.78,150.19
+Effective Latency (CO-aware) (ms),5093.02,591.34,32479.47,4410554.86,837.90,1002.53,1131.32,1605.24,2637.58,6076.77,14553.76,18011.44,22164.45,5486.47
+Effective Prefill Concurrency,0.16,0.00,5.00,,,,,,0.00,,1.00,1.00,2.00,0.46
+Effective Prefill Throughput (tokens/sec),8213.03,0.00,207266.85,,,,,,0.00,,36499.00,72268.58,115366.19,24846.12
+Effective Prefill Throughput Per User (tokens/sec/user),6943.03,0.00,132845.67,,,,,,0.00,,28669.97,59942.41,104735.17,21145.52
+Effective Total Throughput (tokens/sec),8316.29,0.00,207339.55,,,,,,96.03,,36499.00,72500.89,115440.94,24857.70
+HTTP Blocked (ms),0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00
+HTTP Chunks Received,164.05,10.00,1016.00,142066.00,19.00,26.00,30.00,42.00,70.00,200.75,508.00,611.75,771.00,195.59
+HTTP Chunks Sent,1.00,1.00,1.00,866.00,1.00,1.00,1.00,1.00,1.00,1.00,1.00,1.00,1.00,0.00
+HTTP Connecting (ms),0.47,0.00,4.69,405.00,0.00,0.00,0.00,0.00,0.00,0.85,1.07,1.32,3.56,0.64
+HTTP Connection Overhead (ms),0.47,0.00,4.69,405.00,0.00,0.00,0.00,0.00,0.00,0.85,1.07,1.32,3.56,0.64
+HTTP Connection Reused (ratio),0.54,0.00,1.00,466.00,0.00,0.00,0.00,0.00,1.00,1.00,1.00,1.00,1.00,0.50
+HTTP Data Received (KB),43.63,3.05,250.38,37781.34,5.32,7.46,8.60,11.77,19.13,53.41,139.26,156.97,207.41,50.98
+HTTP Data Sent (KB),106.49,36.80,230.43,92220.92,38.13,43.46,48.35,70.46,96.44,138.92,178.39,195.17,218.36,47.05
+HTTP DNS Lookup (ms),0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00
+HTTP Duration (excl. conn) (ms),5091.14,589.78,32477.62,4408930.40,836.58,1000.91,1129.60,1603.30,2635.78,6075.06,14551.37,18008.73,22162.77,5486.42
+HTTP Receiving (ms),4606.31,113.08,31681.73,3989063.00,437.42,618.34,725.74,1110.77,2133.32,5544.88,13706.15,17477.09,21630.04,5471.15
+HTTP Sending (ms),0.79,0.17,6.07,683.06,0.19,0.20,0.21,0.24,0.35,1.16,1.52,1.93,4.31,0.77
+HTTP Total Time (ms),5091.61,591.27,32477.62,4409335.40,836.58,1000.91,1130.13,1603.62,2635.78,6075.06,14552.31,18008.95,22163.55,5486.41
+HTTP Waiting (TTFB) (ms),484.05,306.05,2225.06,419184.34,316.51,329.56,340.26,365.85,400.73,570.94,685.40,803.74,1430.34,200.21
+Input Sequence Length (tokens),26096.19,8094.00,57151.00,22599304.00,8386.15,9848.75,11404.50,16875.00,23735.50,34239.75,44485.00,48870.25,53923.55,11914.06
+Inter Chunk Latency (ms),26.14,0.00,1571.32,3967539.40,0.00,0.00,0.13,18.80,22.02,25.32,41.33,58.07,275.43,39.28
+Inter Token Latency (ms),14.09,3.11,39.09,12200.22,8.94,9.77,10.16,11.25,13.33,15.79,18.58,21.62,27.07,3.86
+OSL Mismatch Diff % (%),-73.39,-98.56,0.00,-63559.30,-96.72,-95.42,-94.55,-92.58,-88.26,-69.30,-10.13,0.00,0.00,30.01
+Output Sequence Length (tokens),329.09,20.00,1958.00,284993.00,38.00,53.00,62.00,85.00,139.00,403.75,1072.50,1205.75,1594.85,394.96
+Output Token Count (tokens),329.09,20.00,1958.00,284993.00,38.00,53.00,62.00,85.00,139.00,403.75,1072.50,1205.75,1594.85,394.96
+Output Token Throughput Per User (tokens/sec/user),75.61,25.58,321.57,65481.83,36.94,46.25,53.82,63.33,75.03,88.85,98.42,102.39,111.91,19.14
+Prefill Throughput Per User (tokens/sec/user),57244.61,5725.08,132845.72,49573831.84,9921.64,17915.08,23971.96,35900.45,52727.09,75869.21,98658.98,110297.41,123485.76,27969.38
+Request Latency (ms),5079.11,589.93,32477.77,4398506.62,829.72,988.58,1112.30,1582.81,2623.38,6062.01,14544.16,17997.79,22162.90,5489.93
+Time to First Output Token (ms),2099.82,321.86,20081.61,1692456.59,344.91,365.06,387.47,434.35,728.17,1995.69,5898.46,9777.39,15095.14,3194.44
+Time to First Token (ms),497.65,308.41,2226.30,430967.22,318.55,337.49,349.06,374.68,413.02,579.97,704.07,898.35,1432.47,207.07
+Time to Second Token (ms),26.78,0.00,585.31,23192.47,0.00,0.00,0.09,14.57,20.61,25.47,41.39,56.96,318.73,47.82
+Tokens In Flight (tokens),48327.53,0.00,315948.39,,,,,,32663.75,,137438.40,169427.95,214713.30,56792.35
+Usage Completion Tokens (tokens),329.09,20.00,1958.00,284993.00,38.00,53.00,62.00,85.00,139.00,403.75,1072.50,1205.75,1594.85,394.96
+Usage Prompt Cache Read Tokens (tokens),25485.13,7744.00,56256.00,21866240.00,7744.00,8886.40,10668.80,16448.00,23168.00,33520.00,43942.40,48048.00,53367.04,11955.15
+Usage Prompt Tokens (tokens),26096.19,8094.00,57151.00,22599304.00,8386.15,9848.75,11404.50,16875.00,23735.50,34239.75,44485.00,48870.25,53923.55,11914.06
+Usage Total Tokens (tokens),26425.29,8180.00,58714.00,22884297.00,8446.15,9936.50,11550.50,17025.25,24041.00,34584.50,44919.50,49466.00,54515.65,12043.69
+
+Metric,Value
+Benchmark Duration (sec),2751.63
+Input Token Throughput (tokens/sec),8213.03
+OSL Mismatch Count,786.00
+Output Token Throughput (tokens/sec),103.57
+Overall Usage Prompt Cache Read % (%),96.76
+Replay Schedule Degraded,0.00
+Replay Schedule Lag p50 (ms),3.00
+Replay Schedule Lag p90 (ms),4.70
+Replay Schedule Lag p99 (ms),6.12
+Request Count,866.00
+Request Throughput (requests/sec),0.31
+Total Input Sequence Length (tokens),22599304.00
+Total Output Sequence Length (tokens),284993.00
+Total Output Tokens (tokens),284993.00
+Total Token Throughput (tokens/sec),8316.60
+Total Usage Completion Tokens (tokens),284993.00
+Total Usage Prompt Cache Read Tokens (tokens),21866240.00
+Total Usage Prompt Tokens (tokens),22599304.00
+Total Usage Total Tokens (tokens),22884297.00
+
+
+Endpoint,GPU_Index,GPU_Name,GPU_UUID,Metric,avg,min,max,sum,p1,p5,p10,p25,p50,p75,p90,p95,p99,std

experiments/kv-cache-fp8/capacity/paced/N0024/profile_export_aiperf.json

@@ -0,0 +1,922 @@
+{
+  "schema_version": "1.4",
+  "aiperf_version": "0.11.0",
+  "benchmark_id": "e2e492e4572f",
+  "request_throughput": {
+    "unit": "requests/sec",
+    "avg": 0.3147213091337059
+  },
+  "request_latency": {
+    "unit": "ms",
+    "avg": 5079.106953927251,
+    "p1": 829.7188436,
+    "p5": 988.5767452499999,
+    "p10": 1112.2956705,
+    "p25": 1582.8125687499999,
+    "p50": 2623.3824295,
+    "p75": 6062.01321225,
+    "p90": 14544.162478999999,
+    "p95": 17997.7850355,
+    "p99": 22162.895364699998,
+    "min": 589.927913,
+    "max": 32477.768427,
+    "std": 5489.930295630577,
+    "count": 866,
+    "sum": 4398506.622101
+  },
+  "request_count": {
+    "unit": "requests",
+    "avg": 866.0
+  },
+  "time_to_first_token": {
+    "unit": "ms",
+    "avg": 497.6526809549653,
+    "p1": 318.5495471,
+    "p5": 337.4857545,
+    "p10": 349.0611055,
+    "p25": 374.679585,
+    "p50": 413.0213725,
+    "p75": 579.97255475,
+    "p90": 704.070852,
+    "p95": 898.34956775,
+    "p99": 1432.4688358000005,
+    "min": 308.410849,
+    "max": 2226.301559,
+    "std": 207.06819539567772,
+    "count": 866,
+    "sum": 430967.221707
+  },
+  "time_to_second_token": {
+    "unit": "ms",
+    "avg": 26.781140304849885,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.08678899999999999,
+    "p25": 14.567177749999999,
+    "p50": 20.6060655,
+    "p75": 25.47079775,
+    "p90": 41.3875825,
+    "p95": 56.959046,
+    "p99": 318.7259600500001,
+    "min": 0.0,
+    "max": 585.3135139999999,
+    "std": 47.8203458863395,
+    "count": 866,
+    "sum": 23192.467504
+  },
+  "inter_token_latency": {
+    "unit": "ms",
+    "avg": 14.088008915135369,
+    "p1": 8.935641015655884,
+    "p5": 9.766988386444861,
+    "p10": 10.160849308398056,
+    "p25": 11.254930828125,
+    "p50": 13.327766593588235,
+    "p75": 15.789976043980069,
+    "p90": 18.581272623666038,
+    "p95": 21.622604528112447,
+    "p99": 27.074647638233266,
+    "min": 3.109699861111111,
+    "max": 39.09466487179487,
+    "std": 3.86199028855266,
+    "count": 866,
+    "sum": 12200.215720507229
+  },
+  "output_token_throughput": {
+    "unit": "tokens/sec",
+    "avg": 103.57202084750836
+  },
+  "output_token_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 75.61412314686162,
+    "p1": 36.93544787800233,
+    "p5": 46.24790322631378,
+    "p10": 53.81807729760632,
+    "p25": 63.33132537284305,
+    "p50": 75.03132615647972,
+    "p75": 88.8499471815648,
+    "p90": 98.41697608307919,
+    "p95": 102.38572930169303,
+    "p99": 111.91195821897855,
+    "min": 25.57893777269484,
+    "max": 321.57444276396984,
+    "std": 19.138412833548262,
+    "count": 866,
+    "sum": 65481.830645182155
+  },
+  "output_sequence_length": {
+    "unit": "tokens",
+    "avg": 329.09122401847577,
+    "p1": 38.0,
+    "p5": 53.0,
+    "p10": 62.0,
+    "p25": 85.0,
+    "p50": 139.0,
+    "p75": 403.75,
+    "p90": 1072.5,
+    "p95": 1205.75,
+    "p99": 1594.8500000000004,
+    "min": 20.0,
+    "max": 1958.0,
+    "std": 394.9637978459474,
+    "count": 866,
+    "sum": 284993.0
+  },
+  "input_sequence_length": {
+    "unit": "tokens",
+    "avg": 26096.19399538106,
+    "p1": 8386.15,
+    "p5": 9848.75,
+    "p10": 11404.5,
+    "p25": 16875.0,
+    "p50": 23735.5,
+    "p75": 34239.75,
+    "p90": 44485.0,
+    "p95": 48870.25,
+    "p99": 53923.55000000002,
+    "min": 8094.0,
+    "max": 57151.0,
+    "std": 11914.062900167835,
+    "count": 866,
+    "sum": 22599304.0
+  },
+  "output_token_count": {
+    "unit": "tokens",
+    "avg": 329.09122401847577,
+    "p1": 38.0,
+    "p5": 53.0,
+    "p10": 62.0,
+    "p25": 85.0,
+    "p50": 139.0,
+    "p75": 403.75,
+    "p90": 1072.5,
+    "p95": 1205.75,
+    "p99": 1594.8500000000004,
+    "min": 20.0,
+    "max": 1958.0,
+    "std": 394.9637978459474,
+    "count": 866,
+    "sum": 284993.0
+  },
+  "inter_chunk_latency": {
+    "unit": "ms",
+    "avg": 26.140928350479324,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.13480499999999998,
+    "p25": 18.8030965,
+    "p50": 22.015335,
+    "p75": 25.3213835,
+    "p90": 41.331579999999995,
+    "p95": 58.073425299999975,
+    "p99": 275.43367066000076,
+    "min": 0.0,
+    "max": 1571.3233309999998,
+    "std": 39.283272571897236,
+    "count": 151775,
+    "sum": 3967539.4003939996
+  },
+  "total_output_tokens": {
+    "unit": "tokens",
+    "avg": 284993.0
+  },
+  "benchmark_duration": {
+    "unit": "sec",
+    "avg": 2751.627440128
+  },
+  "total_isl": {
+    "unit": "tokens",
+    "avg": 22599304.0
+  },
+  "total_osl": {
+    "unit": "tokens",
+    "avg": 284993.0
+  },
+  "telemetry_data": {
+    "summary": {
+      "endpoints_configured": [],
+      "endpoints_successful": [],
+      "start_time": "2026-07-20T21:32:19.701616",
+      "end_time": "2026-07-20T22:18:14.002079"
+    },
+    "endpoints": {},
+    "error_summary": []
+  },
+  "input_config": {
+    "models": {
+      "items": [
+        {
+          "name": "glm-5.2"
+        }
+      ]
+    },
+    "endpoint": {
+      "urls": [
+        "http://127.0.0.1:8391"
+      ],
+      "type": "chat",
+      "api_key": "<redacted>",
+      "streaming": true,
+      "use_server_token_count": true,
+      "extra": {
+        "cache_salt": "2aa386a91b47e21885111b7efc90adba"
+      }
+    },
+    "datasets": [
+      {
+        "name": "main",
+        "type": "file",
+        "path": "results/capacity/sweep-paced-jul20/N0024/paced.jsonl",
+        "format": "dag_jsonl",
+        "entries": 1000000,
+        "random_seed": 0
+      }
+    ],
+    "phases": [
+      {
+        "name": "profiling",
+        "type": "fixed_schedule",
+        "requests": 1000000
+      }
+    ],
+    "artifacts": {
+      "dir": "/home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0024",
+      "records": [
+        "jsonl"
+      ],
+      "raw": true,
+      "trace": true,
+      "slice_duration": 60.0,
+      "auto_plot": false,
+      "plot_required": false
+    },
+    "tokenizer": {
+      "name": "builtin"
+    },
+    "gpu_telemetry": {
+      "enabled": true
+    },
+    "server_metrics": {
+      "enabled": true,
+      "urls": [],
+      "formats": [
+        "json",
+        "csv",
+        "jsonl"
+      ]
+    },
+    "network_latency": {
+      "enabled": false
+    },
+    "otel": {},
+    "mlflow": {},
+    "wandb": {},
+    "runtime": {
+      "ui": "simple"
+    }
+  },
+  "run_info": {
+    "benchmark_id": "e2e492e4572f",
+    "random_seed": 0,
+    "trial": 0,
+    "cli_command": "aiperf profile --model 'glm-5.2' --url 'http://127.0.0.1:8391' --api-key '<redacted>' --endpoint-type 'chat' --streaming --tokenizer 'builtin' --custom-dataset-type 'dag_jsonl' --input-file 'results/capacity/sweep-paced-jul20/N0024/paced.jsonl' --extra-inputs '{\"cache_salt\": \"2aa386a91b47e21885111b7efc90adba\"}' --fixed-schedule --request-count 1000000 --random-seed 0 --use-server-token-count --artifact-dir 'results/capacity/sweep-paced-jul20/N0024' --export-level 'raw' --export-http-trace --slice-duration 60 --server-metrics-formats 'json' 'csv' 'jsonl' --ui 'simple'"
+  },
+  "was_cancelled": false,
+  "error_summary": [],
+  "start_time": "2026-07-20T21:32:19.701616",
+  "end_time": "2026-07-20T22:18:11.342432",
+  "http_req_waiting": {
+    "unit": "ms",
+    "avg": 484.0465816489607,
+    "p1": 316.50604505,
+    "p5": 329.560229,
+    "p10": 340.256078,
+    "p25": 365.847338,
+    "p50": 400.73337899999996,
+    "p75": 570.9416944999999,
+    "p90": 685.402105,
+    "p95": 803.74074275,
+    "p99": 1430.3392115000004,
+    "min": 306.046479,
+    "max": 2225.055023,
+    "std": 200.20565273430333,
+    "count": 866,
+    "sum": 419184.33970799996
+  },
+  "http_req_dns_lookup": {
+    "unit": "ms",
+    "avg": 0.0,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 0.0,
+    "p75": 0.0,
+    "p90": 0.0,
+    "p95": 0.0,
+    "p99": 0.0,
+    "min": 0.0,
+    "max": 0.0,
+    "std": 0.0,
+    "count": 866,
+    "sum": 0.0
+  },
+  "usage_prompt_tokens": {
+    "unit": "tokens",
+    "avg": 26096.19399538106,
+    "p1": 8386.15,
+    "p5": 9848.75,
+    "p10": 11404.5,
+    "p25": 16875.0,
+    "p50": 23735.5,
+    "p75": 34239.75,
+    "p90": 44485.0,
+    "p95": 48870.25,
+    "p99": 53923.55000000002,
+    "min": 8094.0,
+    "max": 57151.0,
+    "std": 11914.062900167835,
+    "count": 866,
+    "sum": 22599304.0
+  },
+  "http_req_chunks_sent": {
+    "unit": "count",
+    "avg": 1.0,
+    "p1": 1.0,
+    "p5": 1.0,
+    "p10": 1.0,
+    "p25": 1.0,
+    "p50": 1.0,
+    "p75": 1.0,
+    "p90": 1.0,
+    "p95": 1.0,
+    "p99": 1.0,
+    "min": 1.0,
+    "max": 1.0,
+    "std": 0.0,
+    "count": 866,
+    "sum": 866.0
+  },
+  "usage_total_tokens": {
+    "unit": "tokens",
+    "avg": 26425.28521939954,
+    "p1": 8446.15,
+    "p5": 9936.5,
+    "p10": 11550.5,
+    "p25": 17025.25,
+    "p50": 24041.0,
+    "p75": 34584.5,
+    "p90": 44919.5,
+    "p95": 49466.0,
+    "p99": 54515.65,
+    "min": 8180.0,
+    "max": 58714.0,
+    "std": 12043.685916991888,
+    "count": 866,
+    "sum": 22884297.0
+  },
+  "http_req_connection_reused": {
+    "unit": "ratio",
+    "avg": 0.5381062355658198,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 1.0,
+    "p75": 1.0,
+    "p90": 1.0,
+    "p95": 1.0,
+    "p99": 1.0,
+    "min": 0.0,
+    "max": 1.0,
+    "std": 0.4985458001136929,
+    "count": 866,
+    "sum": 466.0
+  },
+  "http_req_sending": {
+    "unit": "ms",
+    "avg": 0.7887505311778291,
+    "p1": 0.18907965,
+    "p5": 0.20161474999999998,
+    "p10": 0.21264899999999998,
+    "p25": 0.24075275,
+    "p50": 0.35222749999999997,
+    "p75": 1.15867375,
+    "p90": 1.5173364999999999,
+    "p95": 1.9317992499999999,
+    "p99": 4.309275450000014,
+    "min": 0.165097,
+    "max": 6.073760999999999,
+    "std": 0.7672719470171311,
+    "count": 866,
+    "sum": 683.05796
+  },
+  "http_req_connecting": {
+    "unit": "ms",
+    "avg": 0.46766661316397223,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 0.0,
+    "p75": 0.846416,
+    "p90": 1.071364,
+    "p95": 1.31886675,
+    "p99": 3.5556613500000003,
+    "min": 0.0,
+    "max": 4.692132,
+    "std": 0.6388165137524112,
+    "count": 866,
+    "sum": 404.999287
+  },
+  "http_req_chunks_received": {
+    "unit": "count",
+    "avg": 164.04849884526558,
+    "p1": 19.0,
+    "p5": 26.0,
+    "p10": 30.0,
+    "p25": 42.0,
+    "p50": 70.0,
+    "p75": 200.75,
+    "p90": 508.0,
+    "p95": 611.75,
+    "p99": 771.0,
+    "min": 10.0,
+    "max": 1016.0,
+    "std": 195.59381948469198,
+    "count": 866,
+    "sum": 142066.0
+  },
+  "http_req_receiving": {
+    "unit": "ms",
+    "avg": 4606.308318384526,
+    "p1": 437.42265719999995,
+    "p5": 618.3441312499999,
+    "p10": 725.7434675,
+    "p25": 1110.774979,
+    "p50": 2133.319462,
+    "p75": 5544.877630999999,
+    "p90": 13706.1514575,
+    "p95": 17477.09242925,
+    "p99": 21630.043190600005,
+    "min": 113.079138,
+    "max": 31681.725641999998,
+    "std": 5471.147255716913,
+    "count": 866,
+    "sum": 3989063.0037209997
+  },
+  "http_req_data_sent": {
+    "unit": "KB",
+    "avg": 106.49066401739319,
+    "p1": 38.131640625,
+    "p5": 43.4599609375,
+    "p10": 48.3505859375,
+    "p25": 70.45703125,
+    "p50": 96.435546875,
+    "p75": 138.916259765625,
+    "p90": 178.39404296875,
+    "p95": 195.16943359375,
+    "p99": 218.357763671875,
+    "min": 36.796875,
+    "max": 230.43359375,
+    "std": 47.05155594215007,
+    "count": 866,
+    "sum": 92220.9150390625
+  },
+  "http_req_duration": {
+    "unit": "ms",
+    "avg": 5091.143650564664,
+    "p1": 836.5790806499999,
+    "p5": 1000.9093467499999,
+    "p10": 1129.599078,
+    "p25": 1603.302806,
+    "p50": 2635.7773795,
+    "p75": 6075.058577,
+    "p90": 14551.3741395,
+    "p95": 18008.72546725,
+    "p99": 22162.766215850002,
+    "min": 589.77999,
+    "max": 32477.615608,
+    "std": 5486.418556045309,
+    "count": 866,
+    "sum": 4408930.401389
+  },
+  "http_req_blocked": {
+    "unit": "ms",
+    "avg": 0.0,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 0.0,
+    "p75": 0.0,
+    "p90": 0.0,
+    "p95": 0.0,
+    "p99": 0.0,
+    "min": 0.0,
+    "max": 0.0,
+    "std": 0.0,
+    "count": 866,
+    "sum": 0.0
+  },
+  "http_req_data_received": {
+    "unit": "KB",
+    "avg": 43.62741321449192,
+    "p1": 5.322021484375,
+    "p5": 7.46240234375,
+    "p10": 8.595703125,
+    "p25": 11.768798828125,
+    "p50": 19.12890625,
+    "p75": 53.407958984375,
+    "p90": 139.26318359375,
+    "p95": 156.974853515625,
+    "p99": 207.41401367187504,
+    "min": 3.052734375,
+    "max": 250.3828125,
+    "std": 50.97813693855941,
+    "count": 866,
+    "sum": 37781.33984375
+  },
+  "usage_completion_tokens": {
+    "unit": "tokens",
+    "avg": 329.09122401847577,
+    "p1": 38.0,
+    "p5": 53.0,
+    "p10": 62.0,
+    "p25": 85.0,
+    "p50": 139.0,
+    "p75": 403.75,
+    "p90": 1072.5,
+    "p95": 1205.75,
+    "p99": 1594.8500000000004,
+    "min": 20.0,
+    "max": 1958.0,
+    "std": 394.9637978459474,
+    "count": 866,
+    "sum": 284993.0
+  },
+  "time_to_first_output_token": {
+    "unit": "ms",
+    "avg": 2099.822075239454,
+    "p1": 344.90936085,
+    "p5": 365.0568265,
+    "p10": 387.4706495,
+    "p25": 434.34826599999997,
+    "p50": 728.1698494999999,
+    "p75": 1995.6904217499998,
+    "p90": 5898.458261,
+    "p95": 9777.3914445,
+    "p99": 15095.138197200004,
+    "min": 321.863185,
+    "max": 20081.612343999997,
+    "std": 3194.4437340405457,
+    "count": 806,
+    "sum": 1692456.592643
+  },
+  "osl_mismatch_diff_pct": {
+    "unit": "%",
+    "avg": -73.39411372779618,
+    "p1": -96.71856516008744,
+    "p5": -95.41694404363238,
+    "p10": -94.54965241686763,
+    "p25": -92.58366758538736,
+    "p50": -88.26203614392941,
+    "p75": -69.30388568280696,
+    "p90": -10.13233106256362,
+    "p95": 0.0,
+    "p99": 0.0,
+    "min": -98.55907780979827,
+    "max": 0.0,
+    "std": 30.00706086303778,
+    "count": 866,
+    "sum": -63559.30248827149
+  },
+  "prefill_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 57244.60951748117,
+    "p1": 9921.636674416095,
+    "p5": 17915.077036875133,
+    "p10": 23971.96157287409,
+    "p25": 35900.449303745096,
+    "p50": 52727.08556687091,
+    "p75": 75869.20691013716,
+    "p90": 98658.97706906323,
+    "p95": 110297.41444113538,
+    "p99": 123485.75649428267,
+    "min": 5725.084906949856,
+    "max": 132845.72438550988,
+    "std": 27969.37980394619,
+    "count": 866,
+    "sum": 49573831.84213869
+  },
+  "e2e_output_token_throughput": {
+    "unit": "tokens/sec/user",
+    "avg": 59.65274509427621,
+    "p1": 26.594617104161554,
+    "p5": 35.58858140260494,
+    "p10": 41.169722567624866,
+    "p25": 50.32854438988647,
+    "p50": 59.41119326719533,
+    "p75": 68.6538401167002,
+    "p90": 77.95233257351244,
+    "p95": 84.91435958992713,
+    "p99": 94.8673521125692,
+    "min": 21.119003605409894,
+    "max": 97.80154383357723,
+    "std": 14.347531692810591,
+    "count": 866,
+    "sum": 51659.2772516432
+  },
+  "http_req_connection_overhead": {
+    "unit": "ms",
+    "avg": 0.46766661316397223,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 0.0,
+    "p75": 0.846416,
+    "p90": 1.071364,
+    "p95": 1.31886675,
+    "p99": 3.5556613500000003,
+    "min": 0.0,
+    "max": 4.692132,
+    "std": 0.6388165137524112,
+    "count": 866,
+    "sum": 404.999287
+  },
+  "http_req_total": {
+    "unit": "ms",
+    "avg": 5091.611317177829,
+    "p1": 836.5790806499999,
+    "p5": 1000.9093467499999,
+    "p10": 1130.1299474999998,
+    "p25": 1603.6156695,
+    "p50": 2635.7773795,
+    "p75": 6075.058577,
+    "p90": 14552.3065505,
+    "p95": 18008.9526865,
+    "p99": 22163.5532119,
+    "min": 591.2684489999999,
+    "max": 32477.615608,
+    "std": 5486.414151901096,
+    "count": 866,
+    "sum": 4409335.400676
+  },
+  "osl_mismatch_count": {
+    "unit": "requests",
+    "avg": 786.0
+  },
+  "usage_prompt_cache_read_tokens": {
+    "unit": "tokens",
+    "avg": 25485.128205128207,
+    "p1": 7744.0,
+    "p5": 8886.4,
+    "p10": 10668.8,
+    "p25": 16448.0,
+    "p50": 23168.0,
+    "p75": 33520.0,
+    "p90": 43942.40000000001,
+    "p95": 48047.99999999999,
+    "p99": 53367.03999999999,
+    "min": 7744.0,
+    "max": 56256.0,
+    "std": 11955.150080993877,
+    "count": 858,
+    "sum": 21866240.0
+  },
+  "total_usage_prompt_tokens": {
+    "unit": "tokens",
+    "avg": 22599304.0
+  },
+  "total_usage_prompt_cache_read_tokens": {
+    "unit": "tokens",
+    "avg": 21866240.0
+  },
+  "total_usage_total_tokens": {
+    "unit": "tokens",
+    "avg": 22884297.0
+  },
+  "total_usage_completion_tokens": {
+    "unit": "tokens",
+    "avg": 284993.0
+  },
+  "overall_usage_prompt_cache_read_pct": {
+    "unit": "%",
+    "avg": 96.75625408640903
+  },
+  "input_token_throughput": {
+    "unit": "tokens/sec",
+    "avg": 8213.028337633483
+  },
+  "total_token_throughput": {
+    "unit": "tokens/sec",
+    "avg": 8316.600358480991
+  },
+  "effective_concurrency": {
+    "unit": "requests",
+    "avg": 1.6023404674808042,
+    "p50": 1.0,
+    "p90": 4.0,
+    "p95": 5.0,
+    "p99": 7.0,
+    "min": 0.0,
+    "max": 10.0,
+    "std": 1.859511551550837
+  },
+  "effective_decode_throughput": {
+    "unit": "tokens/sec",
+    "avg": 103.25721867334026,
+    "p50": 37.22240403385088,
+    "p90": 273.92960121662367,
+    "p95": 347.13843863039205,
+    "p99": 517.7051202394446,
+    "min": 0.0,
+    "max": 167077.166033097,
+    "std": 425.94192005307224
+  },
+  "effective_prefill_throughput": {
+    "unit": "tokens/sec",
+    "avg": 8213.028337651394,
+    "p50": 0.0,
+    "p90": 36499.00241490118,
+    "p95": 72268.5777051791,
+    "p99": 115366.1869326599,
+    "min": 0.0,
+    "max": 207266.84514195437,
+    "std": 24846.123003453424
+  },
+  "effective_decode_concurrency": {
+    "unit": "requests",
+    "avg": 1.4457185635659084,
+    "p50": 1.0,
+    "p90": 4.0,
+    "p95": 5.0,
+    "p99": 7.0,
+    "min": 0.0,
+    "max": 9.0,
+    "std": 1.7279744681457418
+  },
+  "effective_prefill_concurrency": {
+    "unit": "requests",
+    "avg": 0.1566219039148957,
+    "p50": 0.0,
+    "p90": 1.0,
+    "p95": 1.0,
+    "p99": 2.0,
+    "min": 0.0,
+    "max": 5.0,
+    "std": 0.46269506096093493
+  },
+  "effective_total_throughput": {
+    "unit": "tokens/sec",
+    "avg": 8316.285637189967,
+    "p50": 96.0298359856486,
+    "p90": 36499.00241490118,
+    "p95": 72500.89434624171,
+    "p99": 115440.94391180859,
+    "min": 0.0,
+    "max": 207339.54744525507,
+    "std": 24857.701458939984
+  },
+  "effective_decode_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 46.04643463382741,
+    "p50": 22.591930631954888,
+    "p90": 112.37361066210624,
+    "p95": 116.39628508014142,
+    "p99": 138.77772346292863,
+    "min": 0.0,
+    "max": 62572.92582681159,
+    "std": 150.1856211704817
+  },
+  "effective_prefill_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 6943.032017111011,
+    "p50": 0.0,
+    "p90": 28669.971559216054,
+    "p95": 59942.40808658679,
+    "p99": 104735.16708234504,
+    "min": 0.0,
+    "max": 132845.67291223595,
+    "std": 21145.52407558143
+  },
+  "tokens_in_flight": {
+    "unit": "tokens",
+    "avg": 48327.52742085344,
+    "p50": 32663.754716948155,
+    "p90": 137438.40171923963,
+    "p95": 169427.94664341898,
+    "p99": 214713.2985797812,
+    "min": 0.0,
+    "max": 315948.39279789967,
+    "std": 56792.34824224233
+  },
+  "active_decode_throughput": {
+    "unit": "tokens/sec",
+    "avg": 173.03714273128423,
+    "p50": 115.183018251071,
+    "p90": 334.7735308239149,
+    "p95": 405.74152271705395,
+    "p99": 608.6621128217685,
+    "min": 0.0,
+    "max": 167077.166033097,
+    "std": 540.3311434147522
+  },
+  "active_prefill_throughput": {
+    "unit": "tokens/sec",
+    "avg": 66496.69301615353,
+    "p50": 63387.1725226273,
+    "p90": 111319.84680489686,
+    "p95": 122990.85302565324,
+    "p99": 158832.64521676823,
+    "min": 5725.0849108564835,
+    "max": 207266.84514195437,
+    "std": 33504.22604571842
+  },
+  "active_decode_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 77.16403408318779,
+    "p50": 78.9489737668446,
+    "p90": 115.52322577414552,
+    "p95": 118.96817048000578,
+    "p99": 201.54812450435057,
+    "min": 0.0,
+    "max": 62572.92582681159,
+    "std": 188.14205479961012
+  },
+  "active_prefill_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 56214.181865976745,
+    "p50": 51569.075130486126,
+    "p90": 100532.0676641673,
+    "p95": 112473.49758452109,
+    "p99": 124404.9043695445,
+    "min": 5725.0849108564835,
+    "max": 132845.67291223595,
+    "std": 29162.845059082545
+  },
+  "active_total_throughput": {
+    "unit": "tokens/sec",
+    "avg": 13386.742356652223,
+    "p50": 179.04988026507718,
+    "p90": 63445.819204965825,
+    "p95": 88316.30687033146,
+    "p99": 123062.39228899337,
+    "min": 37.037590271072084,
+    "max": 207339.54744525507,
+    "std": 30442.841344758494
+  },
+  "credit_to_start_latency": {
+    "unit": "ms",
+    "avg": 1.7198963879907623,
+    "p1": 0.9975424,
+    "p5": 1.113472,
+    "p10": 1.1764480000000002,
+    "p25": 1.2913919999999999,
+    "p50": 1.4670079999999999,
+    "p75": 1.72,
+    "p90": 2.422016,
+    "p95": 3.6204799999999997,
+    "p99": 5.658124800000001,
+    "min": 0.681472,
+    "max": 8.5888,
+    "std": 0.8698983326904065,
+    "count": 866,
+    "sum": 1489.430272
+  },
+  "effective_latency": {
+    "unit": "ms",
+    "avg": 5093.019471519631,
+    "p1": 837.9047296,
+    "p5": 1002.528832,
+    "p10": 1131.317376,
+    "p25": 1605.236672,
+    "p50": 2637.583488,
+    "p75": 6076.772416,
+    "p90": 14553.758848000001,
+    "p95": 18011.441344,
+    "p99": 22164.4454144,
+    "min": 591.339264,
+    "max": 32479.47392,
+    "std": 5486.469207143681,
+    "count": 866,
+    "sum": 4410554.862336
+  },
+  "replay_sched_lag_p50": {
+    "unit": "ms",
+    "avg": 2.9951999999999996
+  },
+  "replay_sched_lag_p90": {
+    "unit": "ms",
+    "avg": 4.6966784
+  },
+  "replay_sched_lag_p99": {
+    "unit": "ms",
+    "avg": 6.12023808
+  },
+  "replay_sched_degraded": {
+    "unit": "count",
+    "avg": 0.0
+  }
+}

experiments/kv-cache-fp8/capacity/paced/N0024/profile_export_console.txt

@@ -0,0 +1,87 @@
+
+
+                                             NVIDIA AIPerf | LLM Metrics: Effective                                              
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                    Metric ┃       avg ┃    min ┃        max ┃        p99 ┃        p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│          Effective Concurrency (requests) │      1.60 │   0.00 │      10.00 │       7.00 │       4.00 │      1.00 │      1.86 │
+│  Effective Decode Throughput (tokens/sec) │    103.26 │   0.00 │ 167,077.17 │     517.71 │     273.93 │     37.22 │    425.94 │
+│ Effective Prefill Throughput (tokens/sec) │  8,213.03 │   0.00 │ 207,266.85 │ 115,366.19 │  36,499.00 │      0.00 │ 24,846.12 │
+│   Effective Decode Concurrency (requests) │      1.45 │   0.00 │       9.00 │       7.00 │       4.00 │      1.00 │      1.73 │
+│  Effective Prefill Concurrency (requests) │      0.16 │   0.00 │       5.00 │       2.00 │       1.00 │      0.00 │      0.46 │
+│   Effective Total Throughput (tokens/sec) │  8,316.29 │   0.00 │ 207,339.55 │ 115,440.94 │  36,499.00 │     96.03 │ 24,857.70 │
+│      Effective Decode Throughput Per User │     46.05 │   0.00 │  62,572.93 │     138.78 │     112.37 │     22.59 │    150.19 │
+│                         (tokens/sec/user) │           │        │            │            │            │           │           │
+│     Effective Prefill Throughput Per User │  6,943.03 │   0.00 │ 132,845.67 │ 104,735.17 │  28,669.97 │      0.00 │ 21,145.52 │
+│                         (tokens/sec/user) │           │        │            │            │            │           │           │
+│                 Tokens In Flight (tokens) │ 48,327.53 │   0.00 │ 315,948.39 │ 214,713.30 │ 137,438.40 │ 32,663.75 │ 56,792.35 │
+│         Effective Latency (CO-aware) (ms) │  5,093.02 │ 591.34 │  32,479.47 │  22,164.45 │  14,553.76 │  2,637.58 │  5,486.47 │
+└───────────────────────────────────────────┴───────────┴────────┴────────────┴────────────┴────────────┴───────────┴───────────┘
+                                              NVIDIA AIPerf | LLM Metrics: Active                                               
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                 Metric ┃       avg ┃      min ┃        max ┃        p99 ┃        p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│  Active Decode Throughput (tokens/sec) │    173.04 │     0.00 │ 167,077.17 │     608.66 │     334.77 │    115.18 │    540.33 │
+│ Active Prefill Throughput (tokens/sec) │ 66,496.69 │ 5,725.08 │ 207,266.85 │ 158,832.65 │ 111,319.85 │ 63,387.17 │ 33,504.23 │
+│      Active Decode Throughput Per User │     77.16 │     0.00 │  62,572.93 │     201.55 │     115.52 │     78.95 │    188.14 │
+│                      (tokens/sec/user) │           │          │            │            │            │           │           │
+│     Active Prefill Throughput Per User │ 56,214.18 │ 5,725.08 │ 132,845.67 │ 124,404.90 │ 100,532.07 │ 51,569.08 │ 29,162.85 │
+│                      (tokens/sec/user) │           │          │            │            │            │           │           │
+│   Active Total Throughput (tokens/sec) │ 13,386.74 │    37.04 │ 207,339.55 │ 123,062.39 │  63,445.82 │    179.05 │ 30,442.84 │
+└────────────────────────────────────────┴───────────┴──────────┴────────────┴────────────┴────────────┴───────────┴───────────┘
+                                                NVIDIA AIPerf | LLM Metrics: Usage                                                
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                  Metric ┃           avg ┃      min ┃       max ┃       p99 ┃       p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│            Usage Prompt Tokens (tokens) │     26,096.19 │ 8,094.00 │ 57,151.00 │ 53,923.55 │ 44,485.00 │ 23,735.50 │ 11,914.06 │
+│ Usage Prompt Cache Read Tokens (tokens) │     25,485.13 │ 7,744.00 │ 56,256.00 │ 53,367.04 │ 43,942.40 │ 23,168.00 │ 11,955.15 │
+│        Usage Completion Tokens (tokens) │        329.09 │    20.00 │  1,958.00 │  1,594.85 │  1,072.50 │    139.00 │    394.96 │
+│             Usage Total Tokens (tokens) │     26,425.29 │ 8,180.00 │ 58,714.00 │ 54,515.65 │ 44,919.50 │ 24,041.00 │ 12,043.69 │
+│      Total Usage Prompt Tokens (tokens) │ 22,599,304.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│    Total Usage Prompt Cache Read Tokens │ 21,866,240.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                (tokens) │               │          │           │           │           │           │           │
+│       Overall Usage Prompt Cache Read % │         96.76 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                     (%) │               │          │           │           │           │           │           │
+│  Total Usage Completion Tokens (tokens) │    284,993.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│       Total Usage Total Tokens (tokens) │ 22,884,297.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+└─────────────────────────────────────────┴───────────────┴──────────┴───────────┴───────────┴───────────┴───────────┴───────────┘
+                                                    NVIDIA AIPerf | LLM Metrics                                                     
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                        Metric ┃       avg ┃      min ┃       max ┃       p99 ┃       p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│                      Time to First Token (ms) │    497.65 │   308.41 │  2,226.30 │  1,432.47 │    704.07 │    413.02 │    207.07 │
+│                     Time to Second Token (ms) │     26.78 │     0.00 │    585.31 │    318.73 │     41.39 │     20.61 │     47.82 │
+│               Time to First Output Token (ms) │  2,099.82 │   321.86 │ 20,081.61 │ 15,095.14 │  5,898.46 │    728.17 │  3,194.44 │
+│                          Request Latency (ms) │  5,079.11 │   589.93 │ 32,477.77 │ 22,162.90 │ 14,544.16 │  2,623.38 │  5,489.93 │
+│                      Inter Token Latency (ms) │     14.09 │     3.11 │     39.09 │     27.07 │     18.58 │     13.33 │      3.86 │
+│              Output Token Throughput Per User │     75.61 │    25.58 │    321.57 │    111.91 │     98.42 │     75.03 │     19.14 │
+│                             (tokens/sec/user) │           │          │           │           │           │           │           │
+│ E2E Output Token Throughput (tokens/sec/user) │     59.65 │    21.12 │     97.80 │     94.87 │     77.95 │     59.41 │     14.35 │
+│               Output Sequence Length (tokens) │    329.09 │    20.00 │  1,958.00 │  1,594.85 │  1,072.50 │    139.00 │    394.96 │
+│                Input Sequence Length (tokens) │ 26,096.19 │ 8,094.00 │ 57,151.00 │ 53,923.55 │ 44,485.00 │ 23,735.50 │ 11,914.06 │
+│          Output Token Throughput (tokens/sec) │    103.57 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│           Input Token Throughput (tokens/sec) │  8,213.03 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│             Request Throughput (requests/sec) │      0.31 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                      Request Count (requests) │    866.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+└───────────────────────────────────────────────┴───────────┴──────────┴───────────┴───────────┴───────────┴───────────┴───────────┘
+
+
+No GPU telemetry data collected during the benchmarking run.
+
+╭─────────────────────────────── Output Sequence Length Mismatch Warning ───────────────────────────────╮
+│  786 of 866 requests (90.8%) have output length differing from requested by more than the threshold.  │
+│  Threshold (tokens): min(requested x 5%, 50)                                                          │
+│  Average mismatch: -73.4%                                                                             │
+│                                                                                                       │
+│  Why: Server hit EOS token before reaching requested output length.                                   │
+│                                                                                                       │
+│  Fix Options:                                                                                         │
+│    - --extra-inputs ignore_eos:true - Generate until max_tokens (vLLM, TensorRT-LLM)                  │
+│    - --extra-inputs min_tokens:<N> - Set minimum output length (vLLM, TensorRT-LLM, SGLang)           │
+│    - --use-server-token-count - Use server-reported token counts if tokenizer mismatch suspected      │
+│                                                                                                       │
+│  Diagnostics:                                                                                         │
+│    - Review profile_export.jsonl -> osl_mismatch_diff_pct for per-request values                      │
+│    - Adjust: AIPERF_METRICS_OSL_MISMATCH_PCT_THRESHOLD=5                                              │
+│    - Adjust: AIPERF_METRICS_OSL_MISMATCH_MAX_TOKEN_THRESHOLD=50                                       │
+╰───────────────────────────────────────────────────────────────────────────────────────────────────────╯

experiments/kv-cache-fp8/capacity/paced/N0024/summary.json

@@ -0,0 +1,25 @@
+{
+  "sent": 866,
+  "records": 866,
+  "valid": 866,
+  "valid_frac": 1.0,
+  "p25_speed_tok_s": 63.3,
+  "p50_speed_tok_s": 75.0,
+  "p95_ttft_ms": 899.0,
+  "p50_ttft_ms": 412.0,
+  "cache_read_ratio": 0.9676,
+  "max_turn_index": 49,
+  "conversations": 24,
+  "recycled_record_frac": 0.0,
+  "inflight_mean": 2.2,
+  "inflight_p95": 7.0,
+  "foreign_p50": 0.0,
+  "foreign_p95": 3.0,
+  "active_sessions_mean": 7.9,
+  "active_sessions_p95": 20.0,
+  "tiers": {
+    "20tok": true,
+    "60tok": true,
+    "180tok": false
+  }
+}

experiments/kv-cache-fp8/capacity/paced/N0024/verify_cache_salt.log

@@ -0,0 +1 @@
+records_checked=866 mismatches=0

experiments/kv-cache-fp8/capacity/paced/N0024/verify_paced.log

@@ -0,0 +1,8 @@
+PASS 866 profiling records
+PASS identity: replayed conversations == scheduled sessions (0 unscheduled, 0 missing)
+PASS identity: 0 conversations ran more than one pass
+PASS identity: 0 conversations with turns != exactly 0..n-1
+PASS schedule: max turn-0 start deviation 6ms (tol 500ms) across 24 sessions
+PASS pacing: 0/842 continuation gaps outside [declared-50ms, declared+0.5s]
+PASS pinning: raw payload present for every record (0 missing — run AIPerf with --export-level raw)
+PASS pinning: 0 requests whose routed_dp_rank differs from the dataset's declaration (unpinned datasets must carry none)

experiments/kv-cache-fp8/capacity/paced/N0028/aiperf_stdout.log

@@ -0,0 +1,279 @@
+32m │   0.00 │ 225,270.85 │ 132,029.31 │  49,783.30 │    114.30 │ 28,966.93 │
+│      Effective Decode Throughput Per User │     50.36 │   0.00 │  63,446.74 │     148.07 │     113.79 │     36.72 │    154.64 │
+│                         (tokens/sec/user) │           │        │            │            │            │           │           │
+│     Effective Prefill Throughput Per User │  8,787.15 │   0.00 │ 189,313.70 │ 115,241.57 │  40,193.21 │      0.00 │ 24,512.35 │
+│                         (tokens/sec/user) │           │        │            │            │            │           │           │
+│                 Tokens In Flight (tokens) │ 75,134.24 │   0.00 │ 458,481.48 │ 329,272.48 │ 187,254.92 │ 52,788.38 │ 80,644.43 │
+│         Effective Latency (CO-aware) (ms) │  6,162.14 │ 785.24 │  32,740.15 │  24,184.98 │  17,081.76 │  3,214.57 │  6,349.03 │
+└───────────────────────────────────────────┴───────────┴────────┴────────────┴────────────┴────────────┴───────────┴───────────┘
+                                              NVIDIA AIPerf | LLM Metrics: Active                                               
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                 Metric ┃       avg ┃      min ┃        max ┃        p99 ┃        p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│  Active Decode Throughput (tokens/sec) │    201.85 │     0.00 │ 157,124.00 │     744.50 │     417.43 │    118.26 │    587.56 │
+│ Active Prefill Throughput (tokens/sec) │ 70,066.24 │ 4,838.40 │ 225,270.85 │ 180,227.17 │ 119,944.20 │ 65,212.10 │ 37,667.47 │
+│      Active Decode Throughput Per User │     74.05 │     0.00 │  63,446.74 │     199.90 │     115.86 │     72.01 │    182.78 │
+│                      (tokens/sec/user) │           │          │            │            │            │           │           │
+│     Active Prefill Throughput Per User │ 58,637.60 │ 4,838.40 │ 189,313.70 │ 152,617.48 │ 105,373.80 │ 53,359.17 │ 32,961.54 │
+│                      (tokens/sec/user) │           │          │            │            │            │           │           │
+│   Active Total Throughput (tokens/sec) │ 15,162.74 │    46.60 │ 225,270.85 │ 140,868.76 │  67,703.52 │    215.89 │ 33,577.68 │
+└────────────────────────────────────────┴───────────┴──────────┴────────────┴────────────┴────────────┴───────────┴───────────┘
+                                                NVIDIA AIPerf | LLM Metrics: Usage                                                
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                  Metric ┃           avg ┃      min ┃       max ┃       p99 ┃       p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│            Usage Prompt Tokens (tokens) │     28,172.98 │ 8,094.00 │ 89,703.00 │ 77,814.54 │ 49,140.00 │ 24,840.00 │ 14,674.57 │
+│ Usage Prompt Cache Read Tokens (tokens) │     27,494.49 │ 5,824.00 │ 89,600.00 │ 76,736.00 │ 48,384.00 │ 24,128.00 │ 14,654.26 │
+│        Usage Completion Tokens (tokens) │        369.32 │    20.00 │  1,957.00 │  1,606.20 │  1,107.20 │    153.00 │    417.91 │
+│             Usage Total Tokens (tokens) │     28,542.30 │ 8,188.00 │ 90,059.00 │ 78,081.54 │ 49,562.80 │ 25,215.00 │ 14,819.24 │
+│      Total Usage Prompt Tokens (tokens) │ 29,835,189.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│    Total Usage Prompt Cache Read Tokens │ 28,896,704.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                (tokens) │               │          │           │           │           │           │           │
+│       Overall Usage Prompt Cache Read % │         96.85 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                     (%) │               │          │           │           │           │           │           │
+│  Total Usage Completion Tokens (tokens) │    391,106.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│       Total Usage Total Tokens (tokens) │ 30,226,295.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+└─────────────────────────────────────────┴───────────────┴──────────┴───────────┴───────────┴───────────┴───────────┴───────────┘
+                                                    NVIDIA AIPerf | LLM Metrics                                                     
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                        Metric ┃       avg ┃      min ┃       max ┃       p99 ┃       p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│                      Time to First Token (ms) │    516.86 │   310.53 │  1,798.73 │  1,494.32 │    735.24 │    430.58 │    219.58 │
+│                     Time to Second Token (ms) │     31.86 │     0.00 │  1,029.16 │    325.91 │     44.63 │     21.81 │     68.57 │
+│               Time to First Output Token (ms) │  2,555.53 │   324.98 │ 29,704.71 │ 21,140.86 │  6,877.53 │    787.16 │  4,148.08 │
+│                          Request Latency (ms) │  6,147.57 │   783.71 │ 32,738.02 │ 24,168.52 │ 17,080.18 │  3,200.60 │  6,352.09 │
+│                      Inter Token Latency (ms) │     15.57 │     6.27 │     32.89 │     27.86 │     21.01 │     15.03 │      4.19 │
+│              Output Token Throughput Per User │     68.85 │    30.40 │    159.45 │    109.52 │     95.38 │     66.53 │     18.01 │
+│                             (tokens/sec/user) │           │          │           │           │           │           │           │
+│ E2E Output Token Throughput (tokens/sec/user) │     55.90 │    22.94 │     97.62 │     94.86 │     76.53 │     54.23 │     14.67 │
+│               Output Sequence Length (tokens) │    369.32 │    20.00 │  1,957.00 │  1,606.20 │  1,107.20 │    153.00 │    417.91 │
+│                Input Sequence Length (tokens) │ 28,172.98 │ 8,094.00 │ 89,703.00 │ 77,814.54 │ 49,140.00 │ 24,840.00 │ 14,674.57 │
+│          Output Token Throughput (tokens/sec) │    137.64 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│           Input Token Throughput (tokens/sec) │ 10,499.79 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│             Request Throughput (requests/sec) │      0.37 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                      Request Count (requests) │  1,059.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+└───────────────────────────────────────────────┴───────────┴──────────┴───────────┴───────────┴───────────┴───────────┴───────────┘
+
+
+No GPU telemetry data collected during the benchmarking run.
+
+╭──────────────────────────────── Output Sequence Length Mismatch Warning ────────────────────────────────╮
+│  923 of 1,059 requests (87.2%) have output length differing from requested by more than the threshold.  │
+│  Threshold (tokens): min(requested x 5%, 50)                                                            │
+│  Average mismatch: -69.8%                                                                               │
+│                                                                                                         │
+│  Why: Server hit EOS token before reaching requested output length.                                     │
+│                                                                                                         │
+│  Fix Options:                                                                                           │
+│    - --extra-inputs ignore_eos:true - Generate until max_tokens (vLLM, TensorRT-LLM)                    │
+│    - --extra-inputs min_tokens:<N> - Set minimum output length (vLLM, TensorRT-LLM, SGLang)             │
+│    - --use-server-token-count - Use server-reported token counts if tokenizer mismatch suspected        │
+│                                                                                                         │
+│  Diagnostics:                                                                                           │
+│    - Review profile_export.jsonl -> osl_mismatch_diff_pct for per-request values                        │
+│    - Adjust: AIPERF_METRICS_OSL_MISMATCH_PCT_THRESHOLD=5                                                │
+│    - Adjust: AIPERF_METRICS_OSL_MISMATCH_MAX_TOKEN_THRESHOLD=50                                         │
+╰─────────────────────────────────────────────────────────────────────────────────────────────────────────╯
+
+CLI Command: aiperf profile --model 'glm-5.2' --url 'http://127.0.0.1:8391' --api-key '<redacted>' --endpoint-type 'chat' --streaming --tokenizer 'builtin' --custom-dataset-type 
+'dag_jsonl' --input-file 'results/capacity/sweep-paced-jul20/N0028/paced.jsonl' --extra-inputs '{"cache_salt": "bcf18463843c70e1dfa2615636bdfcce"}' --fixed-schedule --request-count 
+1000000 --random-seed 0 --use-server-token-count --artifact-dir 'results/capacity/sweep-paced-jul20/N0028' --export-level 'raw' --export-http-trace --slice-duration 60 
+--server-metrics-formats 'json' 'csv' 'jsonl' --ui 'simple'
+Benchmark Duration: 2841.47 sec
+CSV Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0028/profile_export_aiperf.csv
+JSON Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0028/profile_export_aiperf.json
+Raw Records: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0028/profile_export_raw.jsonl
+Server Metrics CSV Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0028/server_metrics_export.csv
+Server Metrics JSON Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0028/server_metrics_export.json
+Timeslice CSV Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0028/profile_export_aiperf_timeslices.csv
+Timeslice JSON Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0028/profile_export_aiperf_timeslices.json
+Log File: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0028/logs/aiperf.log
+
+
+Profiling: 0/1,059 |          |   0% [00:00<?]
+Profiling: 0/1,059 |          |   0% [00:00<?]
+                                              
+
+Profiling: 11/1,059 |          |   1% [02:02<3:13:46]
+Profiling: 11/1,059 |          |   1% [02:02<3:13:46]
+Profiling: 22/1,059 |▏         |   2% [02:36<1:50:23]
+Profiling: 22/1,059 |▏         |   2% [02:36<1:50:23]
+Profiling: 33/1,059 |▎         |   3% [03:00<1:16:23]
+Profiling: 33/1,059 |▎         |   3% [03:00<1:16:23]
+Profiling: 44/1,059 |▍         |   4% [03:48<1:14:53]
+Profiling: 44/1,059 |▍         |   4% [03:48<1:14:53]
+Profiling: 55/1,059 |▌         |   5% [04:44<1:18:05]
+Profiling: 55/1,059 |▌         |   5% [04:44<1:18:05]
+Profiling: 67/1,059 |▋         |   6% [05:06<1:00:18]
+Profiling: 67/1,059 |▋         |   6% [05:06<1:00:18]
+Profiling: 79/1,059 |▋         |   7% [05:28<49:31]  
+Profiling: 79/1,059 |▋         |   7% [05:28<49:31]
+Profiling: 91/1,059 |▊         |   9% [05:40<38:12]
+Profiling: 91/1,059 |▊         |   9% [05:40<38:12]
+Profiling: 103/1,059 |▉         |  10% [05:54<31:38]
+Profiling: 103/1,059 |▉         |  10% [05:54<31:38]
+Profiling: 114/1,059 |█         |  11% [06:08<27:58]
+Profiling: 114/1,059 |█         |  11% [06:08<27:58]
+Profiling: 125/1,059 |█▏        |  12% [06:28<27:50]
+Profiling: 125/1,059 |█▏        |  12% [06:28<27:50]
+Profiling: 137/1,059 |█▎        |  13% [06:52<28:30]
+Profiling: 137/1,059 |█▎        |  13% [06:52<28:30]
+Profiling: 148/1,059 |█▍        |  14% [07:12<28:00]
+Profiling: 148/1,059 |█▍        |  14% [07:12<28:00]
+Profiling: 159/1,059 |█▌        |  15% [07:40<30:46]
+Profiling: 159/1,059 |█▌        |  15% [07:40<30:46]
+Profiling: 171/1,059 |█▌        |  16% [07:58<27:47]
+Profiling: 171/1,059 |█▌        |  16% [07:58<27:47]
+Profiling: 182/1,059 |█▋        |  17% [08:16<26:25]
+Profiling: 182/1,059 |█▋        |  17% [08:16<26:25]
+Profiling: 193/1,059 |█▊        |  18% [08:32<24:35]
+Profiling: 193/1,059 |█▊        |  18% [08:32<24:35]
+Profiling: 204/1,059 |█▉        |  19% [08:50<23:59]
+Profiling: 204/1,059 |█▉        |  19% [08:50<23:59]
+Profiling: 215/1,059 |██        |  20% [09:08<23:29]
+Profiling: 215/1,059 |██        |  20% [09:08<23:29]
+Profiling: 229/1,059 |██▏       |  22% [09:26<21:14]
+Profiling: 229/1,059 |██▏       |  22% [09:26<21:14]
+Profiling: 240/1,059 |██▎       |  23% [09:50<23:27]
+Profiling: 240/1,059 |██▎       |  23% [09:50<23:27]
+Profiling: 251/1,059 |██▎       |  24% [10:06<22:07]
+Profiling: 251/1,059 |██▎       |  24% [10:06<22:07]
+Profiling: 262/1,059 |██▍       |  25% [10:22<21:05]
+Profiling: 262/1,059 |██▍       |  25% [10:22<21:05]
+Profiling: 275/1,059 |██▌       |  26% [10:36<18:32]
+Profiling: 275/1,059 |██▌       |  26% [10:36<18:32]
+Profiling: 288/1,059 |██▋       |  27% [10:52<17:27]
+Profiling: 288/1,059 |██▋       |  27% [10:52<17:27]
+Profiling: 302/1,059 |██▊       |  29% [11:02<14:27]
+Profiling: 302/1,059 |██▊       |  29% [11:02<14:27]
+Profiling: 316/1,059 |██▉       |  30% [11:12<12:29]
+Profiling: 316/1,059 |██▉       |  30% [11:12<12:29]
+Profiling: 327/1,059 |███       |  31% [11:26<13:09]
+Profiling: 327/1,059 |███       |  31% [11:26<13:09]
+Profiling: 339/1,059 |███▏      |  32% [11:40<13:14]
+Profiling: 339/1,059 |███▏      |  32% [11:40<13:14]
+Profiling: 350/1,059 |███▎      |  33% [11:54<13:35]
+Profiling: 350/1,059 |███▎      |  33% [11:54<13:35]
+Profiling: 362/1,059 |███▍      |  34% [12:06<12:50]
+Profiling: 362/1,059 |███▍      |  34% [12:06<12:50]
+Profiling: 373/1,059 |███▌      |  35% [12:16<12:00]
+Profiling: 373/1,059 |███▌      |  35% [12:16<12:00]
+Profiling: 385/1,059 |███▋      |  36% [12:26<11:03]
+Profiling: 385/1,059 |███▋      |  36% [12:26<11:03]
+Profiling: 396/1,059 |███▋      |  37% [12:36<10:38]
+Profiling: 396/1,059 |███▋      |  37% [12:36<10:38]
+Profiling: 409/1,059 |███▊      |  39% [12:44<09:12]
+Profiling: 409/1,059 |███▊      |  39% [12:44<09:12]
+Profiling: 420/1,059 |███▉      |  40% [13:02<11:24]
+Profiling: 420/1,059 |███▉      |  40% [13:02<11:24]
+Profiling: 434/1,059 |████      |  41% [13:16<10:54]
+Profiling: 434/1,059 |████      |  41% [13:16<10:54]
+Profiling: 446/1,059 |████▏     |  42% [13:30<11:03]
+Profiling: 446/1,059 |████▏     |  42% [13:30<11:03]
+Profiling: 457/1,059 |████▎     |  43% [13:44<11:23]
+Profiling: 457/1,059 |████▎     |  43% [13:44<11:23]
+Profiling: 471/1,059 |████▍     |  44% [14:00<11:09]
+Profiling: 471/1,059 |████▍     |  44% [14:00<11:09]
+Profiling: 483/1,059 |████▌     |  46% [14:16<11:28]
+Profiling: 483/1,059 |████▌     |  46% [14:16<11:28]
+Profiling: 494/1,059 |████▋     |  47% [14:32<11:55]
+Profiling: 494/1,059 |████▋     |  47% [14:32<11:55]
+Profiling: 506/1,059 |████▊     |  48% [14:40<10:00]
+Profiling: 506/1,059 |████▊     |  48% [14:40<10:00]
+Profiling: 518/1,059 |████▉     |  49% [14:52<09:33]
+Profiling: 518/1,059 |████▉     |  49% [14:52<09:33]
+Profiling: 529/1,059 |████▉     |  50% [15:16<12:09]
+Profiling: 529/1,059 |████▉     |  50% [15:16<12:09]
+Profiling: 541/1,059 |█████     |  51% [15:30<11:20]
+Profiling: 541/1,059 |█████     |  51% [15:30<11:20]
+Profiling: 552/1,059 |█████▏    |  52% [15:44<10:59]
+Profiling: 552/1,059 |█████▏    |  52% [15:44<10:59]
+Profiling: 564/1,059 |█████▎    |  53% [15:56<09:58]
+Profiling: 564/1,059 |█████▎    |  53% [15:56<09:58]
+Profiling: 575/1,059 |█████▍    |  54% [16:06<09:03]
+Profiling: 575/1,059 |█████▍    |  54% [16:06<09:03]
+Profiling: 586/1,059 |█████▌    |  55% [16:18<08:46]
+Profiling: 586/1,059 |█████▌    |  55% [16:18<08:46]
+Profiling: 597/1,059 |█████▋    |  56% [16:30<08:31]
+Profiling: 597/1,059 |█████▋    |  56% [16:30<08:31]
+Profiling: 608/1,059 |█████▋    |  57% [16:40<07:52]
+Profiling: 608/1,059 |█████▋    |  57% [16:40<07:52]
+Profiling: 620/1,059 |█████▊    |  59% [16:54<07:56]
+Profiling: 620/1,059 |█████▊    |  59% [16:54<07:56]
+Profiling: 631/1,059 |█████▉    |  60% [17:06<07:45]
+Profiling: 631/1,059 |█████▉    |  60% [17:06<07:45]
+Profiling: 644/1,059 |██████    |  61% [17:18<07:08]
+Profiling: 644/1,059 |██████    |  61% [17:18<07:08]
+Profiling: 655/1,059 |██████▏   |  62% [17:32<07:25]
+Profiling: 655/1,059 |██████▏   |  62% [17:32<07:25]
+Profiling: 666/1,059 |██████▎   |  63% [17:44<07:11]
+Profiling: 666/1,059 |██████▎   |  63% [17:44<07:11]
+Profiling: 682/1,059 |██████▍   |  64% [17:52<05:29]
+Profiling: 682/1,059 |██████▍   |  64% [17:52<05:29]
+Profiling: 694/1,059 |██████▌   |  66% [18:08<06:07]
+Profiling: 694/1,059 |██████▌   |  66% [18:08<06:07]
+Profiling: 706/1,059 |██████▋   |  67% [18:22<06:11]
+Profiling: 706/1,059 |██████▋   |  67% [18:22<06:11]
+Profiling: 718/1,059 |██████▊   |  68% [18:40<06:43]
+Profiling: 718/1,059 |██████▊   |  68% [18:40<06:43]
+Profiling: 731/1,059 |██████▉   |  69% [18:54<06:17]
+Profiling: 731/1,059 |██████▉   |  69% [18:54<06:17]
+Profiling: 744/1,059 |███████   |  70% [19:08<05:55]
+Profiling: 744/1,059 |███████   |  70% [19:08<05:55]
+Profiling: 757/1,059 |███████▏  |  71% [19:28<06:18]
+Profiling: 757/1,059 |███████▏  |  71% [19:28<06:18]
+Profiling: 768/1,059 |███████▎  |  73% [19:48<06:49]
+Profiling: 768/1,059 |███████▎  |  73% [19:48<06:49]
+Profiling: 779/1,059 |███████▎  |  74% [20:00<06:09]
+Profiling: 779/1,059 |███████▎  |  74% [20:00<06:09]
+Profiling: 792/1,059 |███████▍  |  75% [20:12<05:18]
+Profiling: 792/1,059 |███████▍  |  75% [20:12<05:18]
+Profiling: 807/1,059 |███████▌  |  76% [20:34<05:24]
+Profiling: 807/1,059 |███████▌  |  76% [20:34<05:24]
+Profiling: 818/1,059 |███████▋  |  77% [20:52<05:32]
+Profiling: 818/1,059 |███████▋  |  77% [20:52<05:32]
+Profiling: 829/1,059 |███████▊  |  78% [21:02<04:47]
+Profiling: 829/1,059 |███████▊  |  78% [21:02<04:47]
+Profiling: 840/1,059 |███████▉  |  79% [21:18<04:46]
+Profiling: 840/1,059 |███████▉  |  79% [21:18<04:46]
+Profiling: 851/1,059 |████████  |  80% [21:42<05:24]
+Profiling: 851/1,059 |████████  |  80% [21:42<05:24]
+Profiling: 862/1,059 |████████▏ |  81% [22:06<05:42]
+Profiling: 862/1,059 |████████▏ |  81% [22:06<05:42]
+Profiling: 875/1,059 |████████▎ |  83% [22:34<05:45]
+Profiling: 875/1,059 |████████▎ |  83% [22:34<05:45]
+Profiling: 886/1,059 |████████▎ |  84% [23:06<06:15]
+Profiling: 886/1,059 |████████▎ |  84% [23:06<06:15]
+Profiling: 897/1,059 |████████▍ |  85% [23:32<06:00]
+Profiling: 897/1,059 |████████▍ |  85% [23:32<06:00]
+Profiling: 908/1,059 |████████▌ |  86% [23:56<05:34]
+Profiling: 908/1,059 |████████▌ |  86% [23:56<05:34]
+Profiling: 919/1,059 |████████▋ |  87% [24:36<06:08]
+Profiling: 919/1,059 |████████▋ |  87% [24:36<06:08]
+Profiling: 931/1,059 |████████▊ |  88% [25:06<05:31]
+Profiling: 931/1,059 |████████▊ |  88% [25:06<05:31]
+Profiling: 942/1,059 |████████▉ |  89% [25:52<05:57]
+Profiling: 942/1,059 |████████▉ |  89% [25:52<05:57]
+Profiling: 953/1,059 |████████▉ |  90% [26:24<05:19]
+Profiling: 953/1,059 |████████▉ |  90% [26:24<05:19]
+Profiling: 964/1,059 |█████████ |  91% [26:54<04:38]
+Profiling: 964/1,059 |█████████ |  91% [26:54<04:38]
+Profiling: 975/1,059 |█████████▏|  92% [27:28<04:10]
+Profiling: 975/1,059 |█████████▏|  92% [27:28<04:10]
+Profiling: 986/1,059 |█████████▎|  93% [28:40<04:54]
+Profiling: 986/1,059 |█████████▎|  93% [28:40<04:54]
+Profiling: 997/1,059 |█████████▍|  94% [29:36<04:30]
+Profiling: 997/1,059 |█████████▍|  94% [29:36<04:30]
+Profiling: 1,008/1,059 |█████████▌|  95% [31:08<04:43]
+Profiling: 1,008/1,059 |█████████▌|  95% [31:08<04:43]
+Profiling: 1,019/1,059 |█████████▌|  96% [34:34<06:19]
+Profiling: 1,019/1,059 |█████████▌|  96% [34:34<06:19]
+Profiling: 1,030/1,059 |█████████▋|  97% [37:06<05:13]
+Profiling: 1,030/1,059 |█████████▋|  97% [37:06<05:13]
+Profiling: 1,041/1,059 |█████████▊|  98% [40:40<04:01]
+Profiling: 1,041/1,059 |█████████▊|  98% [40:40<04:01]
+Profiling: 1,052/1,059 |█████████▉|  99% [43:56<01:43]
+Profiling: 1,052/1,059 |█████████▉|  99% [43:56<01:43]
+Processing Records: 0/1,059 |          |   0% [00:00<?]
+Processing Records: 1,059/1,059 |██████████| 100% [00:01<00:00]
+Processing Records: 1,059/1,059 |██████████| 100% [00:01<00:00]

experiments/kv-cache-fp8/capacity/paced/N0028/manifest.json

@@ -0,0 +1,109 @@
+{
+  "concurrency": 28,
+  "mode": "paced",
+  "cache_salt": "bcf18463843c70e1dfa2615636bdfcce",
+  "command": [
+    "aiperf",
+    "profile",
+    "--model",
+    "glm-5.2",
+    "--url",
+    "http://127.0.0.1:8391",
+    "--api-key",
+    "<key>",
+    "--endpoint-type",
+    "chat",
+    "--streaming",
+    "--tokenizer",
+    "builtin",
+    "--custom-dataset-type",
+    "dag_jsonl",
+    "--input-file",
+    "results/capacity/sweep-paced-jul20/N0028/paced.jsonl",
+    "--extra-inputs",
+    "{\"cache_salt\": \"bcf18463843c70e1dfa2615636bdfcce\"}",
+    "--fixed-schedule",
+    "--request-count",
+    "1000000",
+    "--random-seed",
+    "0",
+    "--use-server-token-count",
+    "--artifact-dir",
+    "results/capacity/sweep-paced-jul20/N0028",
+    "--export-level",
+    "raw",
+    "--export-http-trace",
+    "--slice-duration",
+    "60",
+    "--server-metrics-formats",
+    "json",
+    "csv",
+    "jsonl",
+    "--ui",
+    "simple"
+  ],
+  "dataset_path": "data/capacity/corpus.jsonl",
+  "signature": {
+    "dataset_sha256": "0327d658a267715aebdab45b67b6e0d324616a0faeec8a75d09b8defcd0b2ac2",
+    "url": "http://127.0.0.1:8391",
+    "model": "glm-5.2",
+    "duration": 900,
+    "warmup": 120,
+    "grace": 300,
+    "seed": 0,
+    "request_ceiling": 1000000,
+    "python": "3.12.3",
+    "aiperf": {
+      "version": "0.11.0",
+      "rev": "73d0524f98f30199ca659a8ab586454a823c887d"
+    },
+    "server": {
+      "version": "0.5.15.post1",
+      "model_path": "/home/jl_fs/models/GLM-5.2-FP8",
+      "served_model_name": "glm-5.2",
+      "tp_size": 8,
+      "dp_size": 8,
+      "enable_cache_report": true,
+      "speculative_algorithm": "EAGLE",
+      "max_running_requests": 48,
+      "context_length": 262144,
+      "kv_cache_dtype": "fp8_e4m3",
+      "mem_fraction_static": 0.872,
+      "scheduler_pids": [
+        82149,
+        82150,
+        82151,
+        82152,
+        82153,
+        82154,
+        82155,
+        82156
+      ],
+      "max_total_num_tokens": 234176,
+      "max_model_len": 262144,
+      "model_id": "glm-5.2"
+    },
+    "evaluator_sha256": "3e48dfede43cbf47244f560547eda95d29c671d4fa23b203a5d220a14edf1c80",
+    "verifier_sha256": "6f84bee84368a72ea1445d0961185a0d5d2fd6ba8f8ec4c3fefac619b46a32aa",
+    "runner_sha256": "aa3e1a0f66bcfac6663b5d2e59a37cae9d7264dda6b4c13130a3edf403e77104",
+    "paced": {
+      "rhythm_sha256": "3dfc2e11cd63aa9b0a5ac61db594a9df320457a3a814e3e183c9e83d33490ffe",
+      "idle_cutoff": 300.0,
+      "pacer_sha256": "c9881fc93b09e8d11a3129e4639a889af1d913661a2c74d86ce59e762bd18558"
+    }
+  },
+  "started_utc": "2026-07-20T22:18:43Z",
+  "loadavg_start": [
+    0.63916015625,
+    0.6142578125,
+    0.65234375
+  ],
+  "status": "ok",
+  "elapsed_s": 2853.0,
+  "loadavg_end": [
+    1.0166015625,
+    0.73583984375,
+    0.64404296875
+  ],
+  "returncode": 0
+}

experiments/kv-cache-fp8/capacity/paced/N0028/paced.manifest.json

@@ -0,0 +1,45 @@
+{
+  "corpus": "data/capacity/corpus.jsonl",
+  "corpus_sha256": "0327d658a267715aebdab45b67b6e0d324616a0faeec8a75d09b8defcd0b2ac2",
+  "rhythm": "data/capacity/rhythm.json",
+  "rhythm_sha256": "3dfc2e11cd63aa9b0a5ac61db594a9df320457a3a814e3e183c9e83d33490ffe",
+  "rhythm_source_sha256": "29b6a19e751ff5230771519aab755f80a0f43a4ba9cf96b72d3a6a437ec99276",
+  "settings": {
+    "developers": 28,
+    "span_s": 900.0,
+    "seed": 0,
+    "idle_cutoff_s": 300.0
+  },
+  "arrival_model": "uniform over [0, span], seeded (traces publish no cross-session wall clock; open-loop by construction)",
+  "population_model": "one shuffled population per seed, N takes the prefix; per-session RNG streams keep a session's schedule/gaps identical across N",
+  "gap_model": "i.i.d. per turn from the conditional empirical distribution (marginal preserved; serial correlation not modeled)",
+  "idle_tail_dropped_frac": 0.041,
+  "python": "3.12.3",
+  "turns_total": 1059,
+  "max_session_turns": 50,
+  "max_session_cap_tokens": 68222,
+  "last_arrival_s": 868.4,
+  "max_session_think_s": 1549.6,
+  "gap_percentiles_s": {
+    "source_conditional": {
+      "p10": 1.211,
+      "p25": 2.208,
+      "p50": 4.561,
+      "p75": 9.037,
+      "p90": 62.579,
+      "p95": 146.055,
+      "p99": 208.6
+    },
+    "sampled": {
+      "p10": 1.231,
+      "p25": 2.245,
+      "p50": 4.546,
+      "p75": 9.224,
+      "p90": 55.349,
+      "p95": 121.329,
+      "p99": 218.112
+    }
+  },
+  "ks_distance_sampled_vs_source": 0.0144,
+  "output_sha256": "651f202525afad3b78dbcdeb083c42850295795cbcd41429d43a2660037bf2b6"
+}

experiments/kv-cache-fp8/capacity/paced/N0028/profile_export_aiperf.csv

@@ -0,0 +1,72 @@
+Metric,avg,min,max,sum,p1,p5,p10,p25,p50,p75,p90,p95,p99,std
+Active Decode Throughput (tokens/sec),201.85,0.00,157124.00,,,,,,118.26,,417.43,510.40,744.50,587.56
+Active Decode Throughput Per User (tokens/sec/user),74.05,0.00,63446.74,,,,,,72.01,,115.86,119.18,199.90,182.78
+Active Prefill Throughput (tokens/sec),70066.24,4838.40,225270.85,,,,,,65212.10,,119944.20,137834.30,180227.17,37667.47
+Active Prefill Throughput Per User (tokens/sec/user),58637.60,4838.40,189313.70,,,,,,53359.17,,105373.80,120272.59,152617.48,32961.54
+Active Total Throughput (tokens/sec),15162.74,46.60,225270.85,,,,,,215.89,,67703.52,95878.49,140868.76,33577.68
+Credit-to-Start Latency (ms),1.74,0.64,8.40,1838.13,0.99,1.14,1.19,1.32,1.50,1.76,2.26,3.54,5.84,0.91
+E2E Output Token Throughput (tokens/sec/user),55.90,22.94,97.62,59203.15,28.61,34.45,38.27,45.71,54.23,64.29,76.53,84.39,94.86,14.67
+Effective Concurrency,2.30,0.00,13.00,,,,,,1.00,,6.00,8.00,10.00,2.53
+Effective Decode Concurrency,2.10,0.00,13.00,,,,,,1.00,,6.00,7.00,10.00,2.38
+Effective Decode Throughput (tokens/sec),137.27,0.00,157124.00,,,,,,60.68,,348.07,459.92,652.54,493.59
+Effective Decode Throughput Per User (tokens/sec/user),50.36,0.00,63446.74,,,,,,36.72,,113.79,117.38,148.07,154.64
+Effective Latency (CO-aware) (ms),6162.14,785.24,32740.15,6525706.77,883.09,1094.43,1298.45,1810.38,3214.57,7558.37,17081.76,20709.27,24184.98,6349.03
+Effective Prefill Concurrency,0.19,0.00,5.00,,,,,,0.00,,1.00,1.00,2.00,0.51
+Effective Prefill Throughput (tokens/sec),10499.79,0.00,225270.85,,,,,,0.00,,49568.03,81804.57,131645.11,28949.19
+Effective Prefill Throughput Per User (tokens/sec/user),8787.15,0.00,189313.70,,,,,,0.00,,40193.21,67616.58,115241.57,24512.35
+Effective Total Throughput (tokens/sec),10637.06,0.00,225270.85,,,,,,114.30,,49783.30,81985.48,132029.31,28966.93
+HTTP Blocked (ms),0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00
+HTTP Chunks Received,186.05,10.00,952.00,197022.00,21.00,26.00,32.00,46.00,81.00,225.00,557.00,608.20,780.68,206.57
+HTTP Chunks Sent,1.00,1.00,1.00,1059.00,1.00,1.00,1.00,1.00,1.00,1.00,1.00,1.00,1.00,0.00
+HTTP Connecting (ms),0.44,0.00,11.32,465.33,0.00,0.00,0.00,0.00,0.00,0.83,1.05,1.24,3.00,0.72
+HTTP Connection Overhead (ms),0.44,0.00,11.32,465.33,0.00,0.00,0.00,0.00,0.00,0.83,1.05,1.24,3.00,0.72
+HTTP Connection Reused (ratio),0.58,0.00,1.00,612.00,0.00,0.00,0.00,0.00,1.00,1.00,1.00,1.00,1.00,0.49
+HTTP Data Received (KB),48.62,3.05,244.03,51488.73,5.83,7.54,8.83,12.67,21.37,59.03,141.89,155.36,205.74,53.54
+HTTP Data Sent (KB),114.38,36.80,363.59,121126.15,38.20,43.82,49.52,71.62,100.63,147.00,195.54,214.46,310.24,57.49
+HTTP DNS Lookup (ms),0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00
+HTTP Duration (excl. conn) (ms),6160.24,783.53,32737.83,6523693.30,881.67,1092.89,1296.89,1808.29,3212.81,7556.63,17080.01,20707.56,24183.31,6349.01
+HTTP Receiving (ms),5657.77,409.87,32212.20,5991575.39,510.61,681.85,867.98,1343.37,2693.98,7123.21,16641.82,20167.42,23755.30,6332.59
+HTTP Sending (ms),0.76,0.16,11.59,805.57,0.19,0.21,0.22,0.25,0.32,1.14,1.46,1.77,3.51,0.85
+HTTP Total Time (ms),6160.68,783.53,32737.83,6524158.63,882.23,1092.97,1296.89,1808.86,3212.81,7556.63,17080.70,20707.66,24183.77,6349.00
+HTTP Waiting (TTFB) (ms),501.71,309.28,1797.76,531312.34,317.31,336.64,350.92,378.38,415.86,601.20,720.02,885.85,1462.08,207.26
+Input Sequence Length (tokens),28172.98,8094.00,89703.00,29835189.00,8414.32,9953.40,11878.40,17386.00,24840.00,36129.50,49140.00,54677.40,77814.54,14674.57
+Inter Chunk Latency (ms),28.84,0.00,1236.23,5962925.07,0.00,0.00,0.63,19.02,23.53,27.25,44.44,60.99,298.36,45.35
+Inter Token Latency (ms),15.57,6.27,32.89,16484.21,9.13,9.94,10.48,12.23,15.03,18.36,21.01,22.92,27.86,4.19
+OSL Mismatch Diff % (%),-69.81,-98.99,0.00,-73929.16,-96.51,-95.40,-94.38,-92.09,-86.49,-62.63,0.00,0.00,0.00,32.58
+Output Sequence Length (tokens),369.32,20.00,1957.00,391106.00,42.00,54.00,63.00,90.00,153.00,444.50,1107.20,1195.00,1606.20,417.91
+Output Token Count (tokens),369.32,20.00,1957.00,391106.00,42.00,54.00,63.00,90.00,153.00,444.50,1107.20,1195.00,1606.20,417.91
+Output Token Throughput Per User (tokens/sec/user),68.85,30.40,159.45,72910.96,35.89,43.64,47.60,54.48,66.53,81.78,95.38,100.59,109.52,18.01
+Prefill Throughput Per User (tokens/sec/user),59370.04,4838.40,189313.77,62872876.96,10817.50,17923.29,23781.63,35177.09,54601.48,77883.52,104159.38,119141.19,147881.17,31712.37
+Request Latency (ms),6147.57,783.71,32738.02,6510279.25,875.17,1081.56,1281.97,1789.77,3200.60,7550.16,17080.18,20707.72,24168.52,6352.09
+Time to First Output Token (ms),2555.53,324.98,29704.71,2535086.05,353.26,379.34,394.97,440.77,787.16,2364.47,6877.53,12146.18,21140.86,4148.08
+Time to First Token (ms),516.86,310.53,1798.73,547354.18,322.94,341.71,357.57,387.47,430.58,615.71,735.24,938.55,1494.32,219.58
+Time to Second Token (ms),31.86,0.00,1029.16,33741.15,0.00,0.00,0.13,14.65,21.81,26.92,44.63,62.08,325.91,68.57
+Tokens In Flight (tokens),75134.24,0.00,458481.48,,,,,,52788.38,,187254.92,243652.80,329272.48,80644.43
+Usage Completion Tokens (tokens),369.32,20.00,1957.00,391106.00,42.00,54.00,63.00,90.00,153.00,444.50,1107.20,1195.00,1606.20,417.91
+Usage Prompt Cache Read Tokens (tokens),27494.49,5824.00,89600.00,28896704.00,7744.00,8896.00,10880.00,16736.00,24128.00,35168.00,48384.00,53792.00,76736.00,14654.26
+Usage Prompt Tokens (tokens),28172.98,8094.00,89703.00,29835189.00,8414.32,9953.40,11878.40,17386.00,24840.00,36129.50,49140.00,54677.40,77814.54,14674.57
+Usage Total Tokens (tokens),28542.30,8188.00,90059.00,30226295.00,8502.48,10067.90,11985.80,17616.00,25215.00,36575.50,49562.80,55407.50,78081.54,14819.24
+
+Metric,Value
+Benchmark Duration (sec),2841.47
+Input Token Throughput (tokens/sec),10499.79
+OSL Mismatch Count,923.00
+Output Token Throughput (tokens/sec),137.64
+Overall Usage Prompt Cache Read % (%),96.85
+Replay Schedule Degraded,0.00
+Replay Schedule Lag p50 (ms),2.23
+Replay Schedule Lag p90 (ms),5.73
+Replay Schedule Lag p99 (ms),6.20
+Request Count,1059.00
+Request Throughput (requests/sec),0.37
+Total Input Sequence Length (tokens),29835189.00
+Total Output Sequence Length (tokens),391106.00
+Total Output Tokens (tokens),391106.00
+Total Token Throughput (tokens/sec),10637.43
+Total Usage Completion Tokens (tokens),391106.00
+Total Usage Prompt Cache Read Tokens (tokens),28896704.00
+Total Usage Prompt Tokens (tokens),29835189.00
+Total Usage Total Tokens (tokens),30226295.00
+
+
+Endpoint,GPU_Index,GPU_Name,GPU_UUID,Metric,avg,min,max,sum,p1,p5,p10,p25,p50,p75,p90,p95,p99,std

experiments/kv-cache-fp8/capacity/paced/N0028/profile_export_aiperf.json

@@ -0,0 +1,922 @@
+{
+  "schema_version": "1.4",
+  "aiperf_version": "0.11.0",
+  "benchmark_id": "711b26b14dd0",
+  "request_throughput": {
+    "unit": "requests/sec",
+    "avg": 0.3726900187798512
+  },
+  "request_latency": {
+    "unit": "ms",
+    "avg": 6147.572471134088,
+    "p1": 875.17313948,
+    "p5": 1081.5605074,
+    "p10": 1281.9695422,
+    "p25": 1789.769044,
+    "p50": 3200.596074,
+    "p75": 7550.1574725,
+    "p90": 17080.1770876,
+    "p95": 20707.715049699993,
+    "p99": 24168.522893380006,
+    "min": 783.705116,
+    "max": 32738.01528,
+    "std": 6352.091526925468,
+    "count": 1059,
+    "sum": 6510279.246931
+  },
+  "request_count": {
+    "unit": "requests",
+    "avg": 1059.0
+  },
+  "time_to_first_token": {
+    "unit": "ms",
+    "avg": 516.8594716411709,
+    "p1": 322.93707477999993,
+    "p5": 341.7074963,
+    "p10": 357.566311,
+    "p25": 387.4729605,
+    "p50": 430.581755,
+    "p75": 615.7140509999999,
+    "p90": 735.2391436,
+    "p95": 938.5481345999998,
+    "p99": 1494.3202637800016,
+    "min": 310.52768599999996,
+    "max": 1798.73452,
+    "std": 219.58298543325446,
+    "count": 1059,
+    "sum": 547354.180468
+  },
+  "time_to_second_token": {
+    "unit": "ms",
+    "avg": 31.861331281397543,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.13074580000000008,
+    "p25": 14.6526315,
+    "p50": 21.811156,
+    "p75": 26.915262499999997,
+    "p90": 44.632537,
+    "p95": 62.07817589999999,
+    "p99": 325.90768406000046,
+    "min": 0.0,
+    "max": 1029.164503,
+    "std": 68.56734647680848,
+    "count": 1059,
+    "sum": 33741.149827
+  },
+  "inter_token_latency": {
+    "unit": "ms",
+    "avg": 15.565828285383207,
+    "p1": 9.1306111493489,
+    "p5": 9.941106625765332,
+    "p10": 10.483965714354067,
+    "p25": 12.228335114345368,
+    "p50": 15.031442795979151,
+    "p75": 18.355005476531677,
+    "p90": 21.006483747966456,
+    "p95": 22.91668190960254,
+    "p99": 27.864464628833822,
+    "min": 6.271600507936508,
+    "max": 32.89364961627907,
+    "std": 4.187167414032779,
+    "count": 1059,
+    "sum": 16484.212154220815
+  },
+  "output_token_throughput": {
+    "unit": "tokens/sec",
+    "avg": 137.64051226148487
+  },
+  "output_token_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 68.84887345633788,
+    "p1": 35.88897921347157,
+    "p5": 43.636341085483814,
+    "p10": 47.604349815150414,
+    "p25": 54.48105380724847,
+    "p50": 66.52721322716245,
+    "p75": 81.77731588280407,
+    "p90": 95.38375385062379,
+    "p95": 100.59244425276856,
+    "p99": 109.52236599174921,
+    "min": 30.40100480382997,
+    "max": 159.44893153422834,
+    "std": 18.00997519822924,
+    "count": 1059,
+    "sum": 72910.95699026182
+  },
+  "output_sequence_length": {
+    "unit": "tokens",
+    "avg": 369.3163361661945,
+    "p1": 42.0,
+    "p5": 54.0,
+    "p10": 63.0,
+    "p25": 90.0,
+    "p50": 153.0,
+    "p75": 444.5,
+    "p90": 1107.2,
+    "p95": 1194.999999999999,
+    "p99": 1606.2000000000007,
+    "min": 20.0,
+    "max": 1957.0,
+    "std": 417.91214160336744,
+    "count": 1059,
+    "sum": 391106.0
+  },
+  "input_sequence_length": {
+    "unit": "tokens",
+    "avg": 28172.98300283286,
+    "p1": 8414.32,
+    "p5": 9953.4,
+    "p10": 11878.4,
+    "p25": 17386.0,
+    "p50": 24840.0,
+    "p75": 36129.5,
+    "p90": 49140.0,
+    "p95": 54677.399999999994,
+    "p99": 77814.54000000002,
+    "min": 8094.0,
+    "max": 89703.0,
+    "std": 14674.570677409385,
+    "count": 1059,
+    "sum": 29835189.0
+  },
+  "output_token_count": {
+    "unit": "tokens",
+    "avg": 369.3163361661945,
+    "p1": 42.0,
+    "p5": 54.0,
+    "p10": 63.0,
+    "p25": 90.0,
+    "p50": 153.0,
+    "p75": 444.5,
+    "p90": 1107.2,
+    "p95": 1194.999999999999,
+    "p99": 1606.2000000000007,
+    "min": 20.0,
+    "max": 1957.0,
+    "std": 417.91214160336744,
+    "count": 1059,
+    "sum": 391106.0
+  },
+  "inter_chunk_latency": {
+    "unit": "ms",
+    "avg": 28.838722947762708,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.631897,
+    "p25": 19.01836225,
+    "p50": 23.531599999999997,
+    "p75": 27.25041125,
+    "p90": 44.43805150000002,
+    "p95": 60.98690704999999,
+    "p99": 298.3581666299999,
+    "min": 0.0,
+    "max": 1236.2282699999998,
+    "std": 45.34555046190227,
+    "count": 206768,
+    "sum": 5962925.066463
+  },
+  "total_output_tokens": {
+    "unit": "tokens",
+    "avg": 391106.0
+  },
+  "benchmark_duration": {
+    "unit": "sec",
+    "avg": 2841.4695938560003
+  },
+  "total_isl": {
+    "unit": "tokens",
+    "avg": 29835189.0
+  },
+  "total_osl": {
+    "unit": "tokens",
+    "avg": 391106.0
+  },
+  "telemetry_data": {
+    "summary": {
+      "endpoints_configured": [],
+      "endpoints_successful": [],
+      "start_time": "2026-07-20T22:18:46.846776",
+      "end_time": "2026-07-20T23:06:11.197371"
+    },
+    "endpoints": {},
+    "error_summary": []
+  },
+  "input_config": {
+    "models": {
+      "items": [
+        {
+          "name": "glm-5.2"
+        }
+      ]
+    },
+    "endpoint": {
+      "urls": [
+        "http://127.0.0.1:8391"
+      ],
+      "type": "chat",
+      "api_key": "<redacted>",
+      "streaming": true,
+      "use_server_token_count": true,
+      "extra": {
+        "cache_salt": "bcf18463843c70e1dfa2615636bdfcce"
+      }
+    },
+    "datasets": [
+      {
+        "name": "main",
+        "type": "file",
+        "path": "results/capacity/sweep-paced-jul20/N0028/paced.jsonl",
+        "format": "dag_jsonl",
+        "entries": 1000000,
+        "random_seed": 0
+      }
+    ],
+    "phases": [
+      {
+        "name": "profiling",
+        "type": "fixed_schedule",
+        "requests": 1000000
+      }
+    ],
+    "artifacts": {
+      "dir": "/home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0028",
+      "records": [
+        "jsonl"
+      ],
+      "raw": true,
+      "trace": true,
+      "slice_duration": 60.0,
+      "auto_plot": false,
+      "plot_required": false
+    },
+    "tokenizer": {
+      "name": "builtin"
+    },
+    "gpu_telemetry": {
+      "enabled": true
+    },
+    "server_metrics": {
+      "enabled": true,
+      "urls": [],
+      "formats": [
+        "json",
+        "csv",
+        "jsonl"
+      ]
+    },
+    "network_latency": {
+      "enabled": false
+    },
+    "otel": {},
+    "mlflow": {},
+    "wandb": {},
+    "runtime": {
+      "ui": "simple"
+    }
+  },
+  "run_info": {
+    "benchmark_id": "711b26b14dd0",
+    "random_seed": 0,
+    "trial": 0,
+    "cli_command": "aiperf profile --model 'glm-5.2' --url 'http://127.0.0.1:8391' --api-key '<redacted>' --endpoint-type 'chat' --streaming --tokenizer 'builtin' --custom-dataset-type 'dag_jsonl' --input-file 'results/capacity/sweep-paced-jul20/N0028/paced.jsonl' --extra-inputs '{\"cache_salt\": \"bcf18463843c70e1dfa2615636bdfcce\"}' --fixed-schedule --request-count 1000000 --random-seed 0 --use-server-token-count --artifact-dir 'results/capacity/sweep-paced-jul20/N0028' --export-level 'raw' --export-http-trace --slice-duration 60 --server-metrics-formats 'json' 'csv' 'jsonl' --ui 'simple'"
+  },
+  "was_cancelled": false,
+  "error_summary": [],
+  "start_time": "2026-07-20T22:18:46.846776",
+  "end_time": "2026-07-20T23:06:08.350295",
+  "http_req_blocked": {
+    "unit": "ms",
+    "avg": 0.0,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 0.0,
+    "p75": 0.0,
+    "p90": 0.0,
+    "p95": 0.0,
+    "p99": 0.0,
+    "min": 0.0,
+    "max": 0.0,
+    "std": 0.0,
+    "count": 1059,
+    "sum": 0.0
+  },
+  "http_req_connecting": {
+    "unit": "ms",
+    "avg": 0.43940312936732767,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 0.0,
+    "p75": 0.834461,
+    "p90": 1.0466932,
+    "p95": 1.2441410999999987,
+    "p99": 2.995530220000001,
+    "min": 0.0,
+    "max": 11.322054,
+    "std": 0.7207938670630636,
+    "count": 1059,
+    "sum": 465.32791399999996
+  },
+  "http_req_sending": {
+    "unit": "ms",
+    "avg": 0.7606933701605287,
+    "p1": 0.1903517,
+    "p5": 0.20704289999999997,
+    "p10": 0.2179382,
+    "p25": 0.24583249999999998,
+    "p50": 0.324496,
+    "p75": 1.1403159999999999,
+    "p90": 1.4606978000000002,
+    "p95": 1.7658166999999996,
+    "p99": 3.5127928800000143,
+    "min": 0.15992,
+    "max": 11.590537,
+    "std": 0.850868353042862,
+    "count": 1059,
+    "sum": 805.5742789999999
+  },
+  "http_req_waiting": {
+    "unit": "ms",
+    "avg": 501.71136648158637,
+    "p1": 317.30970232,
+    "p5": 336.6398337,
+    "p10": 350.92175499999996,
+    "p25": 378.376564,
+    "p50": 415.86272099999996,
+    "p75": 601.1950029999999,
+    "p90": 720.0170794,
+    "p95": 885.8466159999999,
+    "p99": 1462.0779608800008,
+    "min": 309.280054,
+    "max": 1797.759368,
+    "std": 207.2613527338644,
+    "count": 1059,
+    "sum": 531312.3371039999
+  },
+  "http_req_data_sent": {
+    "unit": "KB",
+    "avg": 114.37785683722852,
+    "p1": 38.19765625,
+    "p5": 43.823046875,
+    "p10": 49.521484375,
+    "p25": 71.61962890625,
+    "p50": 100.634765625,
+    "p75": 147.00341796875,
+    "p90": 195.5353515625,
+    "p95": 214.46259765624998,
+    "p99": 310.2404687500001,
+    "min": 36.796875,
+    "max": 363.5927734375,
+    "std": 57.48610457838337,
+    "count": 1059,
+    "sum": 121126.150390625
+  },
+  "usage_total_tokens": {
+    "unit": "tokens",
+    "avg": 28542.299338999055,
+    "p1": 8502.48,
+    "p5": 10067.9,
+    "p10": 11985.800000000001,
+    "p25": 17616.0,
+    "p50": 25215.0,
+    "p75": 36575.5,
+    "p90": 49562.8,
+    "p95": 55407.49999999999,
+    "p99": 78081.54000000002,
+    "min": 8188.0,
+    "max": 90059.0,
+    "std": 14819.242319485034,
+    "count": 1059,
+    "sum": 30226295.0
+  },
+  "usage_prompt_tokens": {
+    "unit": "tokens",
+    "avg": 28172.98300283286,
+    "p1": 8414.32,
+    "p5": 9953.4,
+    "p10": 11878.4,
+    "p25": 17386.0,
+    "p50": 24840.0,
+    "p75": 36129.5,
+    "p90": 49140.0,
+    "p95": 54677.399999999994,
+    "p99": 77814.54000000002,
+    "min": 8094.0,
+    "max": 89703.0,
+    "std": 14674.570677409385,
+    "count": 1059,
+    "sum": 29835189.0
+  },
+  "http_req_chunks_sent": {
+    "unit": "count",
+    "avg": 1.0,
+    "p1": 1.0,
+    "p5": 1.0,
+    "p10": 1.0,
+    "p25": 1.0,
+    "p50": 1.0,
+    "p75": 1.0,
+    "p90": 1.0,
+    "p95": 1.0,
+    "p99": 1.0,
+    "min": 1.0,
+    "max": 1.0,
+    "std": 0.0,
+    "count": 1059,
+    "sum": 1059.0
+  },
+  "http_req_receiving": {
+    "unit": "ms",
+    "avg": 5657.767131309725,
+    "p1": 510.61299754,
+    "p5": 681.8470482,
+    "p10": 867.9759752,
+    "p25": 1343.3695149999999,
+    "p50": 2693.981866,
+    "p75": 7123.2064629999995,
+    "p90": 16641.8233522,
+    "p95": 20167.418146499997,
+    "p99": 23755.29785152001,
+    "min": 409.873196,
+    "max": 32212.200708,
+    "std": 6332.587309140319,
+    "count": 1059,
+    "sum": 5991575.392057
+  },
+  "time_to_first_output_token": {
+    "unit": "ms",
+    "avg": 2555.53029691129,
+    "p1": 353.26453382,
+    "p5": 379.33670374999997,
+    "p10": 394.97062569999997,
+    "p25": 440.7691485,
+    "p50": 787.155409,
+    "p75": 2364.46838475,
+    "p90": 6877.5336099999995,
+    "p95": 12146.183808649992,
+    "p99": 21140.85925842002,
+    "min": 324.983293,
+    "max": 29704.713358999998,
+    "std": 4148.081817398819,
+    "count": 992,
+    "sum": 2535086.054536
+  },
+  "http_req_chunks_received": {
+    "unit": "count",
+    "avg": 186.04532577903683,
+    "p1": 21.0,
+    "p5": 26.0,
+    "p10": 32.0,
+    "p25": 46.0,
+    "p50": 81.0,
+    "p75": 225.0,
+    "p90": 557.0,
+    "p95": 608.1999999999998,
+    "p99": 780.6800000000003,
+    "min": 10.0,
+    "max": 952.0,
+    "std": 206.56777422624356,
+    "count": 1059,
+    "sum": 197022.0
+  },
+  "http_req_dns_lookup": {
+    "unit": "ms",
+    "avg": 0.0,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 0.0,
+    "p75": 0.0,
+    "p90": 0.0,
+    "p95": 0.0,
+    "p99": 0.0,
+    "min": 0.0,
+    "max": 0.0,
+    "std": 0.0,
+    "count": 1059,
+    "sum": 0.0
+  },
+  "http_req_data_received": {
+    "unit": "KB",
+    "avg": 48.620139319228045,
+    "p1": 5.83203125,
+    "p5": 7.542675781250001,
+    "p10": 8.831640625,
+    "p25": 12.6689453125,
+    "p50": 21.3681640625,
+    "p75": 59.03173828125,
+    "p90": 141.8896484375,
+    "p95": 155.35537109375,
+    "p99": 205.73542968750007,
+    "min": 3.052734375,
+    "max": 244.0283203125,
+    "std": 53.54166481227653,
+    "count": 1059,
+    "sum": 51488.7275390625
+  },
+  "http_req_connection_reused": {
+    "unit": "ratio",
+    "avg": 0.5779036827195467,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 1.0,
+    "p75": 1.0,
+    "p90": 1.0,
+    "p95": 1.0,
+    "p99": 1.0,
+    "min": 0.0,
+    "max": 1.0,
+    "std": 0.49389372968152995,
+    "count": 1059,
+    "sum": 612.0
+  },
+  "usage_completion_tokens": {
+    "unit": "tokens",
+    "avg": 369.3163361661945,
+    "p1": 42.0,
+    "p5": 54.0,
+    "p10": 63.0,
+    "p25": 90.0,
+    "p50": 153.0,
+    "p75": 444.5,
+    "p90": 1107.2,
+    "p95": 1194.999999999999,
+    "p99": 1606.2000000000007,
+    "min": 20.0,
+    "max": 1957.0,
+    "std": 417.91214160336744,
+    "count": 1059,
+    "sum": 391106.0
+  },
+  "http_req_duration": {
+    "unit": "ms",
+    "avg": 6160.239191161473,
+    "p1": 881.67411826,
+    "p5": 1092.8908397999999,
+    "p10": 1296.8881106000001,
+    "p25": 1808.2903285,
+    "p50": 3212.813733,
+    "p75": 7556.633704999999,
+    "p90": 17080.008336400002,
+    "p95": 20707.55710629999,
+    "p99": 24183.30762098001,
+    "min": 783.529243,
+    "max": 32737.834808,
+    "std": 6349.006468969554,
+    "count": 1059,
+    "sum": 6523693.30344
+  },
+  "e2e_output_token_throughput": {
+    "unit": "tokens/sec/user",
+    "avg": 55.9047730385589,
+    "p1": 28.61432230108596,
+    "p5": 34.450761119662545,
+    "p10": 38.272810809195974,
+    "p25": 45.70842781647569,
+    "p50": 54.2298369586593,
+    "p75": 64.2860849243676,
+    "p90": 76.526767125716,
+    "p95": 84.38795661082521,
+    "p99": 94.85646527860455,
+    "min": 22.936914562493378,
+    "max": 97.6214184161943,
+    "std": 14.670192652556928,
+    "count": 1059,
+    "sum": 59203.15464783388
+  },
+  "osl_mismatch_diff_pct": {
+    "unit": "%",
+    "avg": -69.81034700814772,
+    "p1": -96.5085021468415,
+    "p5": -95.39681313161977,
+    "p10": -94.38470194276107,
+    "p25": -92.08502170056494,
+    "p50": -86.49398704902868,
+    "p75": -62.627504553734056,
+    "p90": 0.0,
+    "p95": 0.0,
+    "p99": 0.0,
+    "min": -98.99446958270488,
+    "max": 0.0,
+    "std": 32.583119051085895,
+    "count": 1059,
+    "sum": -73929.15748162843
+  },
+  "prefill_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 59370.04434559094,
+    "p1": 10817.50015635344,
+    "p5": 17923.28888045327,
+    "p10": 23781.627028763083,
+    "p25": 35177.08527460439,
+    "p50": 54601.478921472204,
+    "p75": 77883.51672198129,
+    "p90": 104159.38487591658,
+    "p95": 119141.18519523746,
+    "p99": 147881.1747970299,
+    "min": 4838.39964055124,
+    "max": 189313.77461307542,
+    "std": 31712.36598378032,
+    "count": 1059,
+    "sum": 62872876.961980805
+  },
+  "http_req_total": {
+    "unit": "ms",
+    "avg": 6160.67859429084,
+    "p1": 882.23484486,
+    "p5": 1092.9720997,
+    "p10": 1296.8881106000001,
+    "p25": 1808.8645745,
+    "p50": 3212.813733,
+    "p75": 7556.633704999999,
+    "p90": 17080.702529200003,
+    "p95": 20707.657239399992,
+    "p99": 24183.76620030001,
+    "min": 783.529243,
+    "max": 32737.834808,
+    "std": 6348.9962457391075,
+    "count": 1059,
+    "sum": 6524158.6313539995
+  },
+  "http_req_connection_overhead": {
+    "unit": "ms",
+    "avg": 0.43940312936732767,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 0.0,
+    "p75": 0.834461,
+    "p90": 1.0466932,
+    "p95": 1.2441410999999987,
+    "p99": 2.995530220000001,
+    "min": 0.0,
+    "max": 11.322054,
+    "std": 0.7207938670630636,
+    "count": 1059,
+    "sum": 465.32791399999996
+  },
+  "osl_mismatch_count": {
+    "unit": "requests",
+    "avg": 923.0
+  },
+  "usage_prompt_cache_read_tokens": {
+    "unit": "tokens",
+    "avg": 27494.485252140817,
+    "p1": 7744.0,
+    "p5": 8896.0,
+    "p10": 10880.0,
+    "p25": 16736.0,
+    "p50": 24128.0,
+    "p75": 35168.0,
+    "p90": 48384.0,
+    "p95": 53792.0,
+    "p99": 76736.0,
+    "min": 5824.0,
+    "max": 89600.0,
+    "std": 14654.263869765708,
+    "count": 1051,
+    "sum": 28896704.0
+  },
+  "total_usage_total_tokens": {
+    "unit": "tokens",
+    "avg": 30226295.0
+  },
+  "total_usage_prompt_cache_read_tokens": {
+    "unit": "tokens",
+    "avg": 28896704.0
+  },
+  "total_usage_prompt_tokens": {
+    "unit": "tokens",
+    "avg": 29835189.0
+  },
+  "total_usage_completion_tokens": {
+    "unit": "tokens",
+    "avg": 391106.0
+  },
+  "overall_usage_prompt_cache_read_pct": {
+    "unit": "%",
+    "avg": 96.85443588106648
+  },
+  "input_token_throughput": {
+    "unit": "tokens/sec",
+    "avg": 10499.789564410206
+  },
+  "total_token_throughput": {
+    "unit": "tokens/sec",
+    "avg": 10637.430076671691
+  },
+  "effective_concurrency": {
+    "unit": "requests",
+    "avg": 2.29592136227488,
+    "p50": 1.0,
+    "p90": 6.0,
+    "p95": 8.0,
+    "p99": 10.0,
+    "min": 0.0,
+    "max": 13.0,
+    "std": 2.534628850689076
+  },
+  "effective_decode_throughput": {
+    "unit": "tokens/sec",
+    "avg": 137.26785759037332,
+    "p50": 60.68113104836668,
+    "p90": 348.06733255604075,
+    "p95": 459.92024546602215,
+    "p99": 652.5404292543327,
+    "min": 0.0,
+    "max": 157123.99993683933,
+    "std": 493.5877272388157
+  },
+  "effective_prefill_throughput": {
+    "unit": "tokens/sec",
+    "avg": 10499.789564949713,
+    "p50": 0.0,
+    "p90": 49568.02940235155,
+    "p95": 81804.56702067894,
+    "p99": 131645.1054328421,
+    "min": 0.0,
+    "max": 225270.8456005012,
+    "std": 28949.186409167338
+  },
+  "effective_decode_concurrency": {
+    "unit": "requests",
+    "avg": 2.103292996743966,
+    "p50": 1.0,
+    "p90": 6.0,
+    "p95": 7.0,
+    "p99": 10.0,
+    "min": 0.0,
+    "max": 13.0,
+    "std": 2.3838326031750063
+  },
+  "effective_prefill_concurrency": {
+    "unit": "requests",
+    "avg": 0.19262836553091392,
+    "p50": 0.0,
+    "p90": 1.0,
+    "p95": 1.0,
+    "p99": 2.0,
+    "min": 0.0,
+    "max": 5.0,
+    "std": 0.5105789922196532
+  },
+  "effective_total_throughput": {
+    "unit": "tokens/sec",
+    "avg": 10637.057387199457,
+    "p50": 114.30168741800709,
+    "p90": 49783.30016843511,
+    "p95": 81985.483918985,
+    "p99": 132029.30898091983,
+    "min": 0.0,
+    "max": 225270.84560050105,
+    "std": 28966.929620286905
+  },
+  "effective_decode_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 50.35775228192747,
+    "p50": 36.72325999028248,
+    "p90": 113.78706179726689,
+    "p95": 117.37892635334178,
+    "p99": 148.070884267016,
+    "min": 0.0,
+    "max": 63446.73543364781,
+    "std": 154.63755488794024
+  },
+  "effective_prefill_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 8787.148542136827,
+    "p50": 0.0,
+    "p90": 40193.208393523375,
+    "p95": 67616.58000172726,
+    "p99": 115241.56889051865,
+    "min": 0.0,
+    "max": 189313.69528064772,
+    "std": 24512.352487824126
+  },
+  "tokens_in_flight": {
+    "unit": "tokens",
+    "avg": 75134.24378769074,
+    "p50": 52788.37735848948,
+    "p90": 187254.9207731553,
+    "p95": 243652.80014911495,
+    "p99": 329272.47940485453,
+    "min": 0.0,
+    "max": 458481.47797606816,
+    "std": 80644.43390870548
+  },
+  "active_decode_throughput": {
+    "unit": "tokens/sec",
+    "avg": 201.8544431186346,
+    "p50": 118.26303347183699,
+    "p90": 417.4270281449256,
+    "p95": 510.3979659719503,
+    "p99": 744.5013372575216,
+    "min": 0.0,
+    "max": 157123.99993683933,
+    "std": 587.5566188387024
+  },
+  "active_prefill_throughput": {
+    "unit": "tokens/sec",
+    "avg": 70066.23554531948,
+    "p50": 65212.1010544377,
+    "p90": 119944.19600148074,
+    "p95": 137834.30488180724,
+    "p99": 180227.1668777191,
+    "min": 4838.399632395368,
+    "max": 225270.8456005012,
+    "std": 37667.465814176416
+  },
+  "active_decode_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 74.05182996262009,
+    "p50": 72.01339695652187,
+    "p90": 115.85662438975464,
+    "p95": 119.18296778988922,
+    "p99": 199.90179244325134,
+    "min": 0.0,
+    "max": 63446.73543364781,
+    "std": 182.78266349450405
+  },
+  "active_prefill_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 58637.59608862391,
+    "p50": 53359.169557192035,
+    "p90": 105373.80189594867,
+    "p95": 120272.59072753023,
+    "p99": 152617.4759950119,
+    "min": 4838.399632395368,
+    "max": 189313.69528064772,
+    "std": 32961.539161492685
+  },
+  "active_total_throughput": {
+    "unit": "tokens/sec",
+    "avg": 15162.740426375303,
+    "p50": 215.88908815053202,
+    "p90": 67703.51598572811,
+    "p95": 95878.49436304953,
+    "p99": 140868.75650527864,
+    "min": 46.59836889046979,
+    "max": 225270.84560050105,
+    "std": 33577.68180625271
+  },
+  "credit_to_start_latency": {
+    "unit": "ms",
+    "avg": 1.735726655335222,
+    "p1": 0.98948096,
+    "p5": 1.1403775999999999,
+    "p10": 1.1918848000000002,
+    "p25": 1.317632,
+    "p50": 1.504256,
+    "p75": 1.758208,
+    "p90": 2.2564864000000004,
+    "p95": 3.536460799999998,
+    "p99": 5.839093760000034,
+    "min": 0.63744,
+    "max": 8.398336,
+    "std": 0.90726533357369,
+    "count": 1059,
+    "sum": 1838.134528
+  },
+  "effective_latency": {
+    "unit": "ms",
+    "avg": 6162.140477159585,
+    "p1": 883.08997632,
+    "p5": 1094.4321791999998,
+    "p10": 1298.4507904000002,
+    "p25": 1810.3776,
+    "p50": 3214.574336,
+    "p75": 7558.374528,
+    "p90": 17081.763328000005,
+    "p95": 20709.26584319999,
+    "p99": 24184.97547776001,
+    "min": 785.243648,
+    "max": 32740.1536,
+    "std": 6349.030204358803,
+    "count": 1059,
+    "sum": 6525706.765312001
+  },
+  "replay_sched_lag_p50": {
+    "unit": "ms",
+    "avg": 2.234368
+  },
+  "replay_sched_lag_p90": {
+    "unit": "ms",
+    "avg": 5.7334784
+  },
+  "replay_sched_lag_p99": {
+    "unit": "ms",
+    "avg": 6.19544576
+  },
+  "replay_sched_degraded": {
+    "unit": "count",
+    "avg": 0.0
+  }
+}

experiments/kv-cache-fp8/capacity/paced/N0028/profile_export_console.txt

@@ -0,0 +1,87 @@
+
+
+                                             NVIDIA AIPerf | LLM Metrics: Effective                                              
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                    Metric ┃       avg ┃    min ┃        max ┃        p99 ┃        p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│          Effective Concurrency (requests) │      2.30 │   0.00 │      13.00 │      10.00 │       6.00 │      1.00 │      2.53 │
+│  Effective Decode Throughput (tokens/sec) │    137.27 │   0.00 │ 157,124.00 │     652.54 │     348.07 │     60.68 │    493.59 │
+│ Effective Prefill Throughput (tokens/sec) │ 10,499.79 │   0.00 │ 225,270.85 │ 131,645.11 │  49,568.03 │      0.00 │ 28,949.19 │
+│   Effective Decode Concurrency (requests) │      2.10 │   0.00 │      13.00 │      10.00 │       6.00 │      1.00 │      2.38 │
+│  Effective Prefill Concurrency (requests) │      0.19 │   0.00 │       5.00 │       2.00 │       1.00 │      0.00 │      0.51 │
+│   Effective Total Throughput (tokens/sec) │ 10,637.06 │   0.00 │ 225,270.85 │ 132,029.31 │  49,783.30 │    114.30 │ 28,966.93 │
+│      Effective Decode Throughput Per User │     50.36 │   0.00 │  63,446.74 │     148.07 │     113.79 │     36.72 │    154.64 │
+│                         (tokens/sec/user) │           │        │            │            │            │           │           │
+│     Effective Prefill Throughput Per User │  8,787.15 │   0.00 │ 189,313.70 │ 115,241.57 │  40,193.21 │      0.00 │ 24,512.35 │
+│                         (tokens/sec/user) │           │        │            │            │            │           │           │
+│                 Tokens In Flight (tokens) │ 75,134.24 │   0.00 │ 458,481.48 │ 329,272.48 │ 187,254.92 │ 52,788.38 │ 80,644.43 │
+│         Effective Latency (CO-aware) (ms) │  6,162.14 │ 785.24 │  32,740.15 │  24,184.98 │  17,081.76 │  3,214.57 │  6,349.03 │
+└───────────────────────────────────────────┴───────────┴────────┴────────────┴────────────┴────────────┴───────────┴───────────┘
+                                              NVIDIA AIPerf | LLM Metrics: Active                                               
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                 Metric ┃       avg ┃      min ┃        max ┃        p99 ┃        p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│  Active Decode Throughput (tokens/sec) │    201.85 │     0.00 │ 157,124.00 │     744.50 │     417.43 │    118.26 │    587.56 │
+│ Active Prefill Throughput (tokens/sec) │ 70,066.24 │ 4,838.40 │ 225,270.85 │ 180,227.17 │ 119,944.20 │ 65,212.10 │ 37,667.47 │
+│      Active Decode Throughput Per User │     74.05 │     0.00 │  63,446.74 │     199.90 │     115.86 │     72.01 │    182.78 │
+│                      (tokens/sec/user) │           │          │            │            │            │           │           │
+│     Active Prefill Throughput Per User │ 58,637.60 │ 4,838.40 │ 189,313.70 │ 152,617.48 │ 105,373.80 │ 53,359.17 │ 32,961.54 │
+│                      (tokens/sec/user) │           │          │            │            │            │           │           │
+│   Active Total Throughput (tokens/sec) │ 15,162.74 │    46.60 │ 225,270.85 │ 140,868.76 │  67,703.52 │    215.89 │ 33,577.68 │
+└────────────────────────────────────────┴───────────┴──────────┴────────────┴────────────┴────────────┴───────────┴───────────┘
+                                                NVIDIA AIPerf | LLM Metrics: Usage                                                
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                  Metric ┃           avg ┃      min ┃       max ┃       p99 ┃       p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│            Usage Prompt Tokens (tokens) │     28,172.98 │ 8,094.00 │ 89,703.00 │ 77,814.54 │ 49,140.00 │ 24,840.00 │ 14,674.57 │
+│ Usage Prompt Cache Read Tokens (tokens) │     27,494.49 │ 5,824.00 │ 89,600.00 │ 76,736.00 │ 48,384.00 │ 24,128.00 │ 14,654.26 │
+│        Usage Completion Tokens (tokens) │        369.32 │    20.00 │  1,957.00 │  1,606.20 │  1,107.20 │    153.00 │    417.91 │
+│             Usage Total Tokens (tokens) │     28,542.30 │ 8,188.00 │ 90,059.00 │ 78,081.54 │ 49,562.80 │ 25,215.00 │ 14,819.24 │
+│      Total Usage Prompt Tokens (tokens) │ 29,835,189.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│    Total Usage Prompt Cache Read Tokens │ 28,896,704.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                (tokens) │               │          │           │           │           │           │           │
+│       Overall Usage Prompt Cache Read % │         96.85 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                     (%) │               │          │           │           │           │           │           │
+│  Total Usage Completion Tokens (tokens) │    391,106.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│       Total Usage Total Tokens (tokens) │ 30,226,295.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+└─────────────────────────────────────────┴───────────────┴──────────┴───────────┴───────────┴───────────┴───────────┴───────────┘
+                                                    NVIDIA AIPerf | LLM Metrics                                                     
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                        Metric ┃       avg ┃      min ┃       max ┃       p99 ┃       p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│                      Time to First Token (ms) │    516.86 │   310.53 │  1,798.73 │  1,494.32 │    735.24 │    430.58 │    219.58 │
+│                     Time to Second Token (ms) │     31.86 │     0.00 │  1,029.16 │    325.91 │     44.63 │     21.81 │     68.57 │
+│               Time to First Output Token (ms) │  2,555.53 │   324.98 │ 29,704.71 │ 21,140.86 │  6,877.53 │    787.16 │  4,148.08 │
+│                          Request Latency (ms) │  6,147.57 │   783.71 │ 32,738.02 │ 24,168.52 │ 17,080.18 │  3,200.60 │  6,352.09 │
+│                      Inter Token Latency (ms) │     15.57 │     6.27 │     32.89 │     27.86 │     21.01 │     15.03 │      4.19 │
+│              Output Token Throughput Per User │     68.85 │    30.40 │    159.45 │    109.52 │     95.38 │     66.53 │     18.01 │
+│                             (tokens/sec/user) │           │          │           │           │           │           │           │
+│ E2E Output Token Throughput (tokens/sec/user) │     55.90 │    22.94 │     97.62 │     94.86 │     76.53 │     54.23 │     14.67 │
+│               Output Sequence Length (tokens) │    369.32 │    20.00 │  1,957.00 │  1,606.20 │  1,107.20 │    153.00 │    417.91 │
+│                Input Sequence Length (tokens) │ 28,172.98 │ 8,094.00 │ 89,703.00 │ 77,814.54 │ 49,140.00 │ 24,840.00 │ 14,674.57 │
+│          Output Token Throughput (tokens/sec) │    137.64 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│           Input Token Throughput (tokens/sec) │ 10,499.79 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│             Request Throughput (requests/sec) │      0.37 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                      Request Count (requests) │  1,059.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+└───────────────────────────────────────────────┴───────────┴──────────┴───────────┴───────────┴───────────┴───────────┴───────────┘
+
+
+No GPU telemetry data collected during the benchmarking run.
+
+╭──────────────────────────────── Output Sequence Length Mismatch Warning ────────────────────────────────╮
+│  923 of 1,059 requests (87.2%) have output length differing from requested by more than the threshold.  │
+│  Threshold (tokens): min(requested x 5%, 50)                                                            │
+│  Average mismatch: -69.8%                                                                               │
+│                                                                                                         │
+│  Why: Server hit EOS token before reaching requested output length.                                     │
+│                                                                                                         │
+│  Fix Options:                                                                                           │
+│    - --extra-inputs ignore_eos:true - Generate until max_tokens (vLLM, TensorRT-LLM)                    │
+│    - --extra-inputs min_tokens:<N> - Set minimum output length (vLLM, TensorRT-LLM, SGLang)             │
+│    - --use-server-token-count - Use server-reported token counts if tokenizer mismatch suspected        │
+│                                                                                                         │
+│  Diagnostics:                                                                                           │
+│    - Review profile_export.jsonl -> osl_mismatch_diff_pct for per-request values                        │
+│    - Adjust: AIPERF_METRICS_OSL_MISMATCH_PCT_THRESHOLD=5                                                │
+│    - Adjust: AIPERF_METRICS_OSL_MISMATCH_MAX_TOKEN_THRESHOLD=50                                         │
+╰─────────────────────────────────────────────────────────────────────────────────────────────────────────╯

experiments/kv-cache-fp8/capacity/paced/N0028/summary.json

@@ -0,0 +1,25 @@
+{
+  "sent": 1059,
+  "records": 1059,
+  "valid": 1059,
+  "valid_frac": 1.0,
+  "p25_speed_tok_s": 54.5,
+  "p50_speed_tok_s": 66.5,
+  "p95_ttft_ms": 941.0,
+  "p50_ttft_ms": 431.0,
+  "cache_read_ratio": 0.9685,
+  "max_turn_index": 49,
+  "conversations": 28,
+  "recycled_record_frac": 0.0,
+  "inflight_mean": 3.0,
+  "inflight_p95": 9.0,
+  "foreign_p50": 0.0,
+  "foreign_p95": 3.0,
+  "active_sessions_mean": 9.4,
+  "active_sessions_p95": 23.0,
+  "tiers": {
+    "20tok": true,
+    "60tok": false,
+    "180tok": false
+  }
+}

experiments/kv-cache-fp8/capacity/paced/N0028/verify_cache_salt.log

@@ -0,0 +1 @@
+records_checked=1059 mismatches=0

experiments/kv-cache-fp8/capacity/paced/N0028/verify_paced.log

@@ -0,0 +1,8 @@
+PASS 1059 profiling records
+PASS identity: replayed conversations == scheduled sessions (0 unscheduled, 0 missing)
+PASS identity: 0 conversations ran more than one pass
+PASS identity: 0 conversations with turns != exactly 0..n-1
+PASS schedule: max turn-0 start deviation 6ms (tol 500ms) across 28 sessions
+PASS pacing: 0/1031 continuation gaps outside [declared-50ms, declared+0.5s]
+PASS pinning: raw payload present for every record (0 missing — run AIPerf with --export-level raw)
+PASS pinning: 0 requests whose routed_dp_rank differs from the dataset's declaration (unpinned datasets must carry none)

experiments/kv-cache-fp8/capacity/paced/N0032/aiperf_stdout.log

@@ -0,0 +1,273 @@
+━━━━━━━━━━┩
+│  Active Decode Throughput (tokens/sec) │    227.29 │     0.00 │ 139,868.72 │     851.26 │     482.10 │    129.43 │    631.86 │
+│ Active Prefill Throughput (tokens/sec) │ 68,854.35 │ 4,963.69 │ 222,958.93 │ 170,940.13 │ 115,309.75 │ 65,449.12 │ 35,310.41 │
+│      Active Decode Throughput Per User │     69.18 │     0.00 │  48,841.47 │     191.05 │     114.02 │     68.01 │    173.80 │
+│                      (tokens/sec/user) │           │          │            │            │            │           │           │
+│     Active Prefill Throughput Per User │ 56,232.77 │ 4,963.69 │ 174,877.28 │ 133,523.70 │ 103,409.90 │ 50,516.26 │ 31,454.82 │
+│                      (tokens/sec/user) │           │          │            │            │            │           │           │
+│   Active Total Throughput (tokens/sec) │ 17,206.87 │    36.57 │ 223,087.15 │ 133,927.05 │  73,526.17 │    261.82 │ 34,475.70 │
+└────────────────────────────────────────┴───────────┴──────────┴────────────┴────────────┴────────────┴───────────┴───────────┘
+                                                NVIDIA AIPerf | LLM Metrics: Usage                                                
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                  Metric ┃           avg ┃      min ┃       max ┃       p99 ┃       p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│            Usage Prompt Tokens (tokens) │     27,173.74 │ 8,094.00 │ 85,205.00 │ 73,467.30 │ 46,635.00 │ 24,222.00 │ 13,653.18 │
+│ Usage Prompt Cache Read Tokens (tokens) │     26,497.97 │ 5,824.00 │ 85,120.00 │ 65,914.88 │ 46,105.60 │ 23,424.00 │ 13,659.02 │
+│        Usage Completion Tokens (tokens) │        353.41 │    32.00 │  2,265.00 │  1,599.10 │  1,105.00 │    152.00 │    411.11 │
+│             Usage Total Tokens (tokens) │     27,527.15 │ 8,203.00 │ 85,403.00 │ 74,108.80 │ 46,899.00 │ 24,621.00 │ 13,772.09 │
+│      Total Usage Prompt Tokens (tokens) │ 32,907,398.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│    Total Usage Prompt Cache Read Tokens │ 31,877,056.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                (tokens) │               │          │           │           │           │           │           │
+│       Overall Usage Prompt Cache Read % │         96.87 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                     (%) │               │          │           │           │           │           │           │
+│  Total Usage Completion Tokens (tokens) │    427,981.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│       Total Usage Total Tokens (tokens) │ 33,335,379.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+└─────────────────────────────────────────┴───────────────┴──────────┴───────────┴───────────┴───────────┴───────────┴───────────┘
+                                                    NVIDIA AIPerf | LLM Metrics                                                     
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                        Metric ┃       avg ┃      min ┃       max ┃       p99 ┃       p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│                      Time to First Token (ms) │    534.83 │   309.14 │  2,402.46 │  1,574.95 │    791.98 │    433.10 │    247.34 │
+│                     Time to Second Token (ms) │     35.31 │     0.00 │    926.38 │    334.56 │     48.75 │     23.37 │     76.22 │
+│               Time to First Output Token (ms) │  2,533.55 │   336.17 │ 28,296.26 │ 20,166.16 │  7,177.95 │    802.64 │  4,123.09 │
+│                          Request Latency (ms) │  6,357.72 │   681.07 │ 43,595.13 │ 27,560.11 │ 17,796.36 │  3,135.82 │  6,981.83 │
+│                      Inter Token Latency (ms) │     16.44 │     5.53 │     39.56 │     31.30 │     22.13 │     15.87 │      4.73 │
+│              Output Token Throughput Per User │     65.72 │    25.28 │    180.67 │    105.04 │     93.03 │     63.00 │     18.20 │
+│                             (tokens/sec/user) │           │          │           │           │           │           │           │
+│ E2E Output Token Throughput (tokens/sec/user) │     52.95 │    15.91 │     97.25 │     91.22 │     72.88 │     50.97 │     14.35 │
+│               Output Sequence Length (tokens) │    353.41 │    32.00 │  2,265.00 │  1,599.10 │  1,105.00 │    152.00 │    411.11 │
+│                Input Sequence Length (tokens) │ 27,173.74 │ 8,094.00 │ 85,205.00 │ 73,467.30 │ 46,635.00 │ 24,222.00 │ 13,653.18 │
+│          Output Token Throughput (tokens/sec) │    154.92 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│           Input Token Throughput (tokens/sec) │ 11,912.05 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│             Request Throughput (requests/sec) │      0.44 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│          Completed Requests (Success + Error) │  1,213.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                    (requests) │           │          │           │           │           │           │           │
+│                        Request Error Rate (%) │      0.16 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                      Request Count (requests) │  1,211.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│          Time to First Token (error-adjusted) │       N/A │   309.14 │       N/A │  1,592.15 │    794.03 │    433.64 │       N/A │
+│                                          (ms) │           │          │           │           │           │           │           │
+│              Request Latency (error-adjusted) │       N/A │   681.07 │       N/A │ 28,301.96 │ 17,973.89 │  3,146.36 │       N/A │
+│                                          (ms) │           │          │           │           │           │           │           │
+│          Inter Token Latency (error-adjusted) │       N/A │     5.53 │       N/A │     31.47 │     22.26 │     15.88 │       N/A │
+│                                          (ms) │           │          │           │           │           │           │           │
+└───────────────────────────────────────────────┴───────────┴──────────┴───────────┴───────────┴───────────┴───────────┴───────────┘
+
+
+No GPU telemetry data collected during the benchmarking run.
+
+╭───────────────────────────────── Output Sequence Length Mismatch Warning ─────────────────────────────────╮
+│  1,067 of 1,211 requests (88.1%) have output length differing from requested by more than the threshold.  │
+│  Threshold (tokens): min(requested x 5%, 50)                                                              │
+│  Average mismatch: -71.3%                                                                                 │
+│                                                                                                           │
+│  Why: Server hit EOS token before reaching requested output length.                                       │
+│                                                                                                           │
+│  Fix Options:                                                                                             │
+│    - --extra-inputs ignore_eos:true - Generate until max_tokens (vLLM, TensorRT-LLM)                      │
+│    - --extra-inputs min_tokens:<N> - Set minimum output length (vLLM, TensorRT-LLM, SGLang)               │
+│    - --use-server-token-count - Use server-reported token counts if tokenizer mismatch suspected          │
+│                                                                                                           │
+│  Diagnostics:                                                                                             │
+│    - Review profile_export.jsonl -> osl_mismatch_diff_pct for per-request values                          │
+│    - Adjust: AIPERF_METRICS_OSL_MISMATCH_PCT_THRESHOLD=5                                                  │
+│    - Adjust: AIPERF_METRICS_OSL_MISMATCH_MAX_TOKEN_THRESHOLD=50                                           │
+╰───────────────────────────────────────────────────────────────────────────────────────────────────────────╯
+
+CLI Command: aiperf profile --model 'glm-5.2' --url 'http://127.0.0.1:8391' --api-key '<redacted>' --endpoint-type 'chat' --streaming --tokenizer 'builtin' --custom-dataset-type 
+'dag_jsonl' --input-file 'results/capacity/sweep-paced-jul20/N0032/paced.jsonl' --extra-inputs '{"cache_salt": "d513fc35125fa06a944a846295b1a179"}' --fixed-schedule --request-count 
+1000000 --random-seed 0 --use-server-token-count --artifact-dir 'results/capacity/sweep-paced-jul20/N0032' --export-level 'raw' --export-http-trace --slice-duration 60 
+--server-metrics-formats 'json' 'csv' 'jsonl' --ui 'simple'
+Benchmark Duration: 2762.52 sec
+CSV Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0032/profile_export_aiperf.csv
+JSON Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0032/profile_export_aiperf.json
+Raw Records: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0032/profile_export_raw.jsonl
+Server Metrics CSV Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0032/server_metrics_export.csv
+Server Metrics JSON Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0032/server_metrics_export.json
+Timeslice CSV Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0032/profile_export_aiperf_timeslices.csv
+Timeslice JSON Export: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0032/profile_export_aiperf_timeslices.json
+Log File: /home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0032/logs/aiperf.log
+
+
+Profiling: 0/1,213 |          |   0% [00:00<?]
+Profiling: 0/1,213 |          |   0% [00:00<?]
+                                              
+
+Profiling: 14/1,213 |          |   1% [02:10<3:05:36]
+Profiling: 14/1,213 |          |   1% [02:10<3:05:36]
+Profiling: 28/1,213 |▏         |   2% [02:46<1:45:25]
+Profiling: 28/1,213 |▏         |   2% [02:46<1:45:25]
+Profiling: 42/1,213 |▎         |   3% [03:36<1:28:26]
+Profiling: 42/1,213 |▎         |   3% [03:36<1:28:26]
+Profiling: 55/1,213 |▍         |   5% [04:36<1:28:04]
+Profiling: 55/1,213 |▍         |   5% [04:36<1:28:04]
+Profiling: 68/1,213 |▌         |   6% [05:00<1:08:56]
+Profiling: 68/1,213 |▌         |   6% [05:00<1:08:56]
+Profiling: 81/1,213 |▋         |   7% [05:20<55:06]  
+Profiling: 81/1,213 |▋         |   7% [05:20<55:06]
+Profiling: 95/1,213 |▊         |   8% [05:44<46:48]
+Profiling: 95/1,213 |▊         |   8% [05:44<46:48]
+Profiling: 109/1,213 |▉         |   9% [05:58<37:08]
+Profiling: 109/1,213 |▉         |   9% [05:58<37:08]
+Profiling: 123/1,213 |█         |  10% [06:20<34:05]
+Profiling: 123/1,213 |█         |  10% [06:20<34:05]
+Profiling: 136/1,213 |█         |  11% [06:40<31:53]
+Profiling: 136/1,213 |█         |  11% [06:40<31:53]
+Profiling: 149/1,213 |█▏        |  12% [07:04<31:52]
+Profiling: 149/1,213 |█▏        |  12% [07:04<31:52]
+Profiling: 163/1,213 |█▎        |  13% [07:38<34:55]
+Profiling: 163/1,213 |█▎        |  13% [07:38<34:55]
+Profiling: 177/1,213 |█▍        |  15% [08:06<34:29]
+Profiling: 177/1,213 |█▍        |  15% [08:06<34:29]
+Profiling: 190/1,213 |█▌        |  16% [08:20<29:29]
+Profiling: 190/1,213 |█▌        |  16% [08:20<29:29]
+Profiling: 203/1,213 |█▋        |  17% [08:40<28:10]
+Profiling: 203/1,213 |█▋        |  17% [08:40<28:10]
+Profiling: 217/1,213 |█▊        |  18% [09:06<28:43]
+Profiling: 217/1,213 |█▊        |  18% [09:06<28:43]
+Profiling: 232/1,213 |█▉        |  19% [09:26<26:12]
+Profiling: 232/1,213 |█▉        |  19% [09:26<26:12]
+Profiling: 246/1,213 |██        |  20% [09:48<25:40]
+Profiling: 246/1,213 |██        |  20% [09:48<25:40]
+Profiling: 260/1,213 |██▏       |  21% [10:08<24:31]
+Profiling: 260/1,213 |██▏       |  21% [10:08<24:31]
+Profiling: 273/1,213 |██▎       |  23% [10:18<20:43]
+Profiling: 273/1,213 |██▎       |  23% [10:18<20:43]
+Profiling: 286/1,213 |██▎       |  24% [10:34<20:01]
+Profiling: 286/1,213 |██▎       |  24% [10:34<20:01]
+Profiling: 299/1,213 |██▍       |  25% [10:50<19:27]
+Profiling: 299/1,213 |██▍       |  25% [10:50<19:27]
+Profiling: 313/1,213 |██▌       |  26% [11:04<17:52]
+Profiling: 313/1,213 |██▌       |  26% [11:04<17:52]
+Profiling: 326/1,213 |██▋       |  27% [11:16<16:27]
+Profiling: 326/1,213 |██▋       |  27% [11:16<16:27]
+Profiling: 339/1,213 |██▊       |  28% [11:26<14:44]
+Profiling: 339/1,213 |██▊       |  28% [11:26<14:44]
+Profiling: 353/1,213 |██▉       |  29% [11:40<14:27]
+Profiling: 353/1,213 |██▉       |  29% [11:40<14:27]
+Profiling: 369/1,213 |███       |  30% [11:58<14:44]
+Profiling: 369/1,213 |███       |  30% [11:58<14:44]
+Profiling: 382/1,213 |███▏      |  31% [12:14<15:13]
+Profiling: 382/1,213 |███▏      |  31% [12:14<15:13]
+Profiling: 396/1,213 |███▎      |  33% [12:30<15:09]
+Profiling: 396/1,213 |███▎      |  33% [12:30<15:09]
+Profiling: 409/1,213 |███▎      |  34% [12:42<14:11]
+Profiling: 409/1,213 |███▎      |  34% [12:42<14:11]
+Profiling: 422/1,213 |███▍      |  35% [12:54<13:26]
+Profiling: 422/1,213 |███▍      |  35% [12:54<13:26]
+Profiling: 436/1,213 |███▌      |  36% [13:10<13:41]
+Profiling: 436/1,213 |███▌      |  36% [13:10<13:41]
+Profiling: 449/1,213 |███▋      |  37% [13:24<13:32]
+Profiling: 449/1,213 |███▋      |  37% [13:24<13:32]
+Profiling: 463/1,213 |███▊      |  38% [13:32<11:23]
+Profiling: 463/1,213 |███▊      |  38% [13:32<11:23]
+Profiling: 482/1,213 |███▉      |  40% [13:46<10:18]
+Profiling: 482/1,213 |███▉      |  40% [13:46<10:18]
+Profiling: 496/1,213 |████      |  41% [13:58<10:08]
+Profiling: 496/1,213 |████      |  41% [13:58<10:08]
+Profiling: 510/1,213 |████▏     |  42% [14:12<10:27]
+Profiling: 510/1,213 |████▏     |  42% [14:12<10:27]
+Profiling: 525/1,213 |████▎     |  43% [14:22<09:26]
+Profiling: 525/1,213 |████▎     |  43% [14:22<09:26]
+Profiling: 540/1,213 |████▍     |  45% [14:36<09:36]
+Profiling: 540/1,213 |████▍     |  45% [14:36<09:36]
+Profiling: 553/1,213 |████▌     |  46% [14:46<09:09]
+Profiling: 553/1,213 |████▌     |  46% [14:46<09:09]
+Profiling: 566/1,213 |████▋     |  47% [14:58<09:15]
+Profiling: 566/1,213 |████▋     |  47% [14:58<09:15]
+Profiling: 582/1,213 |████▊     |  48% [15:16<09:57]
+Profiling: 582/1,213 |████▊     |  48% [15:16<09:57]
+Profiling: 595/1,213 |████▉     |  49% [15:28<09:41]
+Profiling: 595/1,213 |████▉     |  49% [15:28<09:41]
+Profiling: 609/1,213 |█████     |  50% [15:42<09:38]
+Profiling: 609/1,213 |█████     |  50% [15:42<09:38]
+Profiling: 622/1,213 |█████▏    |  51% [16:02<11:03]
+Profiling: 622/1,213 |█████▏    |  51% [16:02<11:03]
+Profiling: 635/1,213 |█████▏    |  52% [16:18<11:07]
+Profiling: 635/1,213 |█████▏    |  52% [16:18<11:07]
+Profiling: 649/1,213 |█████▎    |  54% [16:34<10:49]
+Profiling: 649/1,213 |█████▎    |  54% [16:34<10:49]
+Profiling: 662/1,213 |█████▍    |  55% [16:44<09:33]
+Profiling: 662/1,213 |█████▍    |  55% [16:44<09:33]
+Profiling: 678/1,213 |█████▌    |  56% [17:00<09:09]
+Profiling: 678/1,213 |█████▌    |  56% [17:00<09:09]
+Profiling: 692/1,213 |█████▋    |  57% [17:10<08:06]
+Profiling: 692/1,213 |█████▋    |  57% [17:10<08:06]
+Profiling: 708/1,213 |█████▊    |  58% [17:22<07:21]
+Profiling: 708/1,213 |█████▊    |  58% [17:22<07:21]
+Profiling: 724/1,213 |█████▉    |  60% [17:36<07:07]
+Profiling: 724/1,213 |█████▉    |  60% [17:36<07:07]
+Profiling: 737/1,213 |██████    |  61% [17:48<07:02]
+Profiling: 737/1,213 |██████    |  61% [17:48<07:02]
+Profiling: 750/1,213 |██████▏   |  62% [18:08<08:14]
+Profiling: 750/1,213 |██████▏   |  62% [18:08<08:14]
+Profiling: 764/1,213 |██████▎   |  63% [18:22<07:50]
+Profiling: 764/1,213 |██████▎   |  63% [18:22<07:50]
+Profiling: 777/1,213 |██████▍   |  64% [18:36<07:40]
+Profiling: 777/1,213 |██████▍   |  64% [18:36<07:40]
+Profiling: 790/1,213 |██████▌   |  65% [18:44<06:32]
+Profiling: 790/1,213 |██████▌   |  65% [18:44<06:32]
+Profiling: 803/1,213 |██████▌   |  66% [19:00<06:57]
+Profiling: 803/1,213 |██████▌   |  66% [19:00<06:57]
+Profiling: 816/1,213 |██████▋   |  67% [19:08<05:56]
+Profiling: 816/1,213 |██████▋   |  67% [19:08<05:56]
+Profiling: 833/1,213 |██████▊   |  69% [19:22<05:31]
+Profiling: 833/1,213 |██████▊   |  69% [19:22<05:31]
+Profiling: 848/1,213 |██████▉   |  70% [19:36<05:25]
+Profiling: 848/1,213 |██████▉   |  70% [19:36<05:25]
+Profiling: 861/1,213 |███████   |  71% [19:50<05:31]
+Profiling: 861/1,213 |███████   |  71% [19:50<05:31]
+Profiling: 874/1,213 |███████▏  |  72% [20:02<05:17]
+Profiling: 874/1,213 |███████▏  |  72% [20:02<05:17]
+Profiling: 888/1,213 |███████▎  |  73% [20:22<05:53]
+Profiling: 888/1,213 |███████▎  |  73% [20:22<05:53]
+Profiling: 901/1,213 |███████▍  |  74% [20:36<05:38]
+Profiling: 901/1,213 |███████▍  |  74% [20:36<05:38]
+Profiling: 915/1,213 |███████▌  |  75% [20:56<05:54]
+Profiling: 915/1,213 |███████▌  |  75% [20:56<05:54]
+Profiling: 928/1,213 |███████▋  |  77% [21:12<05:42]
+Profiling: 928/1,213 |███████▋  |  77% [21:12<05:42]
+Profiling: 941/1,213 |███████▊  |  78% [21:30<05:41]
+Profiling: 941/1,213 |███████▊  |  78% [21:30<05:41]
+Profiling: 954/1,213 |███████▊  |  79% [21:44<05:11]
+Profiling: 954/1,213 |███████▊  |  79% [21:44<05:11]
+Profiling: 967/1,213 |███████▉  |  80% [22:04<05:20]
+Profiling: 967/1,213 |███████▉  |  80% [22:04<05:20]
+Profiling: 980/1,213 |████████  |  81% [22:28<05:41]
+Profiling: 980/1,213 |████████  |  81% [22:28<05:41]
+Profiling: 993/1,213 |████████▏ |  82% [22:52<05:47]
+Profiling: 993/1,213 |████████▏ |  82% [22:52<05:47]
+Profiling: 1,006/1,213 |████████▎ |  83% [23:18<05:52]
+Profiling: 1,006/1,213 |████████▎ |  83% [23:18<05:52]
+Profiling: 1,019/1,213 |████████▍ |  84% [23:38<05:21]
+Profiling: 1,019/1,213 |████████▍ |  84% [23:38<05:21]
+Profiling: 1,032/1,213 |████████▌ |  85% [24:00<05:01]
+Profiling: 1,032/1,213 |████████▌ |  85% [24:00<05:01]
+Profiling: 1,045/1,213 |████████▌ |  86% [24:20<04:33]
+Profiling: 1,045/1,213 |████████▌ |  86% [24:20<04:33]
+Profiling: 1,059/1,213 |████████▋ |  87% [25:16<06:06]
+Profiling: 1,059/1,213 |████████▋ |  87% [25:16<06:06]
+Profiling: 1,073/1,213 |████████▊ |  88% [26:14<06:49]
+Profiling: 1,073/1,213 |████████▊ |  88% [26:14<06:49]
+Profiling: 1,086/1,213 |████████▉ |  90% [26:44<05:48]
+Profiling: 1,086/1,213 |████████▉ |  90% [26:44<05:48]
+Profiling: 1,099/1,213 |█████████ |  91% [27:08<04:42]
+Profiling: 1,099/1,213 |█████████ |  91% [27:08<04:42]
+Profiling: 1,112/1,213 |█████████▏|  92% [27:36<04:00]
+Profiling: 1,112/1,213 |█████████▏|  92% [27:36<04:00]
+Profiling: 1,125/1,213 |█████████▎|  93% [29:02<05:20]
+Profiling: 1,125/1,213 |█████████▎|  93% [29:02<05:20]
+Profiling: 1,138/1,213 |█████████▍|  94% [30:22<05:29]
+Profiling: 1,138/1,213 |█████████▍|  94% [30:22<05:29]
+Profiling: 1,151/1,213 |█████████▍|  95% [31:26<04:42]
+Profiling: 1,151/1,213 |█████████▍|  95% [31:26<04:42]
+Profiling: 1,164/1,213 |█████████▌|  96% [34:18<05:50]
+Profiling: 1,164/1,213 |███████���█▌|  96% [34:18<05:50]
+Profiling: 1,177/1,213 |█████████▋|  97% [35:34<04:03]
+Profiling: 1,177/1,213 |█████████▋|  97% [35:34<04:03]
+Profiling: 1,190/1,213 |█████████▊|  98% [39:24<03:50]
+Profiling: 1,190/1,213 |█████████▊|  98% [39:24<03:50]
+Profiling: 1,203/1,213 |█████████▉|  99% [42:42<01:55]
+Profiling: 1,203/1,213 |█████████▉|  99% [42:42<01:55]
+Processing Records: 0/1,213 |          |   0% [00:00<?]
+Processing Records: 1,213/1,213 |██████████| 100% [00:00<00:00]
+Processing Records: 1,213/1,213 |██████████| 100% [00:00<00:00]

experiments/kv-cache-fp8/capacity/paced/N0032/manifest.json

@@ -0,0 +1,109 @@
+{
+  "concurrency": 32,
+  "mode": "paced",
+  "cache_salt": "d513fc35125fa06a944a846295b1a179",
+  "command": [
+    "aiperf",
+    "profile",
+    "--model",
+    "glm-5.2",
+    "--url",
+    "http://127.0.0.1:8391",
+    "--api-key",
+    "<key>",
+    "--endpoint-type",
+    "chat",
+    "--streaming",
+    "--tokenizer",
+    "builtin",
+    "--custom-dataset-type",
+    "dag_jsonl",
+    "--input-file",
+    "results/capacity/sweep-paced-jul20/N0032/paced.jsonl",
+    "--extra-inputs",
+    "{\"cache_salt\": \"d513fc35125fa06a944a846295b1a179\"}",
+    "--fixed-schedule",
+    "--request-count",
+    "1000000",
+    "--random-seed",
+    "0",
+    "--use-server-token-count",
+    "--artifact-dir",
+    "results/capacity/sweep-paced-jul20/N0032",
+    "--export-level",
+    "raw",
+    "--export-http-trace",
+    "--slice-duration",
+    "60",
+    "--server-metrics-formats",
+    "json",
+    "csv",
+    "jsonl",
+    "--ui",
+    "simple"
+  ],
+  "dataset_path": "data/capacity/corpus.jsonl",
+  "signature": {
+    "dataset_sha256": "0327d658a267715aebdab45b67b6e0d324616a0faeec8a75d09b8defcd0b2ac2",
+    "url": "http://127.0.0.1:8391",
+    "model": "glm-5.2",
+    "duration": 900,
+    "warmup": 120,
+    "grace": 300,
+    "seed": 0,
+    "request_ceiling": 1000000,
+    "python": "3.12.3",
+    "aiperf": {
+      "version": "0.11.0",
+      "rev": "73d0524f98f30199ca659a8ab586454a823c887d"
+    },
+    "server": {
+      "version": "0.5.15.post1",
+      "model_path": "/home/jl_fs/models/GLM-5.2-FP8",
+      "served_model_name": "glm-5.2",
+      "tp_size": 8,
+      "dp_size": 8,
+      "enable_cache_report": true,
+      "speculative_algorithm": "EAGLE",
+      "max_running_requests": 48,
+      "context_length": 262144,
+      "kv_cache_dtype": "fp8_e4m3",
+      "mem_fraction_static": 0.872,
+      "scheduler_pids": [
+        82149,
+        82150,
+        82151,
+        82152,
+        82153,
+        82154,
+        82155,
+        82156
+      ],
+      "max_total_num_tokens": 234176,
+      "max_model_len": 262144,
+      "model_id": "glm-5.2"
+    },
+    "evaluator_sha256": "3e48dfede43cbf47244f560547eda95d29c671d4fa23b203a5d220a14edf1c80",
+    "verifier_sha256": "6f84bee84368a72ea1445d0961185a0d5d2fd6ba8f8ec4c3fefac619b46a32aa",
+    "runner_sha256": "aa3e1a0f66bcfac6663b5d2e59a37cae9d7264dda6b4c13130a3edf403e77104",
+    "paced": {
+      "rhythm_sha256": "3dfc2e11cd63aa9b0a5ac61db594a9df320457a3a814e3e183c9e83d33490ffe",
+      "idle_cutoff": 300.0,
+      "pacer_sha256": "c9881fc93b09e8d11a3129e4639a889af1d913661a2c74d86ce59e762bd18558"
+    }
+  },
+  "started_utc": "2026-07-20T17:08:05Z",
+  "loadavg_start": [
+    0.7646484375,
+    0.87939453125,
+    0.8212890625
+  ],
+  "status": "ok",
+  "elapsed_s": 2773.5,
+  "loadavg_end": [
+    0.2578125,
+    0.5595703125,
+    0.64111328125
+  ],
+  "returncode": 0
+}

experiments/kv-cache-fp8/capacity/paced/N0032/paced.manifest.json

@@ -0,0 +1,45 @@
+{
+  "corpus": "data/capacity/corpus.jsonl",
+  "corpus_sha256": "0327d658a267715aebdab45b67b6e0d324616a0faeec8a75d09b8defcd0b2ac2",
+  "rhythm": "data/capacity/rhythm.json",
+  "rhythm_sha256": "3dfc2e11cd63aa9b0a5ac61db594a9df320457a3a814e3e183c9e83d33490ffe",
+  "rhythm_source_sha256": "29b6a19e751ff5230771519aab755f80a0f43a4ba9cf96b72d3a6a437ec99276",
+  "settings": {
+    "developers": 32,
+    "span_s": 900.0,
+    "seed": 0,
+    "idle_cutoff_s": 300.0
+  },
+  "arrival_model": "uniform over [0, span], seeded (traces publish no cross-session wall clock; open-loop by construction)",
+  "population_model": "one shuffled population per seed, N takes the prefix; per-session RNG streams keep a session's schedule/gaps identical across N",
+  "gap_model": "i.i.d. per turn from the conditional empirical distribution (marginal preserved; serial correlation not modeled)",
+  "idle_tail_dropped_frac": 0.041,
+  "python": "3.12.3",
+  "turns_total": 1213,
+  "max_session_turns": 50,
+  "max_session_cap_tokens": 68222,
+  "last_arrival_s": 883.7,
+  "max_session_think_s": 1549.6,
+  "gap_percentiles_s": {
+    "source_conditional": {
+      "p10": 1.211,
+      "p25": 2.208,
+      "p50": 4.561,
+      "p75": 9.037,
+      "p90": 62.579,
+      "p95": 146.055,
+      "p99": 208.6
+    },
+    "sampled": {
+      "p10": 1.239,
+      "p25": 2.245,
+      "p50": 4.555,
+      "p75": 9.314,
+      "p90": 55.349,
+      "p95": 121.329,
+      "p99": 218.112
+    }
+  },
+  "ks_distance_sampled_vs_source": 0.0148,
+  "output_sha256": "46bc4c22e2541c42e5d4cdc82eea9a7a55bb90189f7df6ab32fa26df3fb8b4b9"
+}

experiments/kv-cache-fp8/capacity/paced/N0032/profile_export_aiperf.csv

@@ -0,0 +1,78 @@
+Metric,avg,min,max,sum,p1,p5,p10,p25,p50,p75,p90,p95,p99,std
+Active Decode Throughput (tokens/sec),227.29,0.00,139868.72,,,,,,129.43,,482.10,562.77,851.26,631.86
+Active Decode Throughput Per User (tokens/sec/user),69.18,0.00,48841.47,,,,,,68.01,,114.02,117.23,191.05,173.80
+Active Prefill Throughput (tokens/sec),68854.35,4963.69,222958.93,,,,,,65449.12,,115309.75,127760.10,170940.13,35310.41
+Active Prefill Throughput Per User (tokens/sec/user),56232.77,4963.69,174877.28,,,,,,50516.26,,103409.90,115483.01,133523.70,31454.82
+Active Total Throughput (tokens/sec),17206.87,36.57,223087.15,,,,,,261.82,,73526.17,99462.72,133927.05,34475.70
+Inter Token Latency (error-adjusted) (ms),,5.53,,,9.51,10.17,10.75,12.75,15.88,19.21,22.26,24.47,31.47,
+Request Latency (error-adjusted) (ms),,681.07,,,890.41,1102.27,1267.25,1879.60,3146.36,7763.19,17973.89,22040.74,28301.96,
+Time to First Token (error-adjusted) (ms),,309.14,,,321.13,343.87,357.53,384.90,433.64,633.74,794.03,984.01,1592.15,
+Credit-to-Start Latency (ms),1.57,0.65,8.06,1898.91,0.95,1.05,1.11,1.22,1.39,1.61,2.02,2.93,4.72,0.70
+E2E Output Token Throughput (tokens/sec/user),52.95,15.91,97.25,64116.42,25.35,32.40,36.13,43.37,50.97,61.32,72.88,80.56,91.22,14.35
+Effective Concurrency,2.79,0.00,15.00,,,,,,2.00,,7.00,9.00,11.00,3.01
+Effective Decode Concurrency,2.56,0.00,15.00,,,,,,2.00,,7.00,8.00,11.00,2.82
+Effective Decode Throughput (tokens/sec),154.49,0.00,139868.72,,,,,,58.82,,420.80,514.04,718.68,531.61
+Effective Decode Throughput Per User (tokens/sec/user),47.02,0.00,48841.47,,,,,,32.91,,109.00,115.44,142.39,146.88
+Effective Latency (CO-aware) (ms),6363.99,3.30,43596.52,7719525.21,909.91,1104.47,1276.41,1898.88,3152.61,7598.24,17795.24,21776.74,27554.61,6977.18
+Effective Prefill Concurrency,0.23,0.00,5.00,,,,,,0.00,,1.00,1.00,3.00,0.58
+Effective Prefill Throughput (tokens/sec),11912.05,0.00,222958.93,,,,,,0.00,,58509.72,85334.27,126391.95,29899.92
+Effective Prefill Throughput Per User (tokens/sec/user),9728.47,0.00,174877.28,,,,,,0.00,,43885.26,70265.52,112919.29,24971.72
+Effective Total Throughput (tokens/sec),12066.54,0.00,223087.15,,,,,,162.31,,58677.51,85553.14,126709.14,29925.39
+HTTP Blocked (ms),0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00
+HTTP Chunks Received,182.26,15.00,1207.00,220711.00,21.00,26.00,31.00,45.00,80.00,221.50,576.00,621.50,795.10,210.70
+HTTP Chunks Sent,1.00,1.00,1.00,1211.00,1.00,1.00,1.00,1.00,1.00,1.00,1.00,1.00,1.00,0.00
+HTTP Connecting (ms),0.39,0.00,5.14,470.46,0.00,0.00,0.00,0.00,0.00,0.78,0.96,1.17,2.28,0.58
+HTTP Connection Overhead (ms),0.39,0.00,5.14,470.46,0.00,0.00,0.00,0.00,0.00,0.78,0.96,1.17,2.28,0.58
+HTTP Connection Reused (ratio),0.59,0.00,1.00,717.00,0.00,0.00,0.00,0.00,1.00,1.00,1.00,1.00,1.00,0.49
+HTTP Data Received (KB),46.90,4.58,292.44,56796.54,6.06,7.34,8.60,12.22,20.97,56.64,143.03,159.80,206.47,53.24
+HTTP Data Sent (KB),111.23,36.80,342.59,134699.12,38.22,43.41,49.08,69.82,98.95,145.88,188.97,207.24,290.35,54.38
+HTTP DNS Lookup (ms),0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00,0.00
+HTTP Duration (excl. conn) (ms),6372.79,680.92,43594.98,7717448.56,912.25,1104.25,1276.03,1901.21,3151.31,7611.29,17796.02,21797.86,27559.97,6978.15
+HTTP Receiving (ms),5859.05,326.60,43209.38,7095309.40,522.89,667.46,831.29,1389.27,2619.28,7137.41,17348.57,21276.23,26847.98,6958.85
+HTTP Sending (ms),0.69,0.17,6.15,835.31,0.19,0.20,0.21,0.23,0.30,1.07,1.33,1.72,3.03,0.70
+HTTP Total Time (ms),6373.18,681.87,43595.90,7717919.02,912.57,1104.59,1277.09,1901.21,3151.31,7611.66,17798.18,21797.86,27559.97,6978.12
+HTTP Waiting (TTFB) (ms),513.05,307.90,2402.11,621303.85,318.67,336.60,349.69,376.53,422.38,620.95,732.59,901.54,1494.82,219.78
+Input Sequence Length (tokens),27173.74,8094.00,85205.00,32907398.00,8445.70,9849.50,11809.00,16810.50,24222.00,35490.00,46635.00,51372.00,73467.30,13653.18
+Inter Chunk Latency (ms),30.90,0.00,1693.72,7051522.77,0.00,0.10,0.81,19.16,24.66,27.82,47.01,62.56,302.72,52.20
+Inter Token Latency (ms),16.44,5.53,39.56,19908.85,9.52,10.16,10.75,12.75,15.87,19.17,22.13,24.39,31.30,4.73
+OSL Mismatch Diff % (%),-71.29,-97.52,0.00,-86336.89,-96.33,-95.52,-94.64,-92.21,-86.84,-64.77,0.00,0.00,0.00,31.55
+Output Sequence Length (tokens),353.41,32.00,2265.00,427981.00,42.00,52.00,61.00,88.00,152.00,432.50,1105.00,1226.50,1599.10,411.11
+Output Token Count (tokens),353.41,32.00,2265.00,427981.00,42.00,52.00,61.00,88.00,152.00,432.50,1105.00,1226.50,1599.10,411.11
+Output Token Throughput Per User (tokens/sec/user),65.72,25.28,180.67,79590.11,31.95,41.00,45.19,52.15,63.00,78.45,93.03,98.43,105.04,18.20
+Prefill Throughput Per User (tokens/sec/user),56624.96,4963.69,174877.33,68572830.76,9392.38,16690.60,22060.36,32265.97,51225.15,75475.53,103016.31,114198.85,130804.34,30523.41
+Request Latency (ms),6357.72,681.07,43595.13,7699201.21,890.57,1101.05,1267.25,1878.88,3135.82,7611.41,17796.36,21797.98,27560.11,6981.83
+Time to First Output Token (ms),2533.55,336.17,28296.26,2842640.57,356.14,378.14,400.04,450.23,802.64,2180.85,7177.95,12000.74,20166.16,4123.09
+Time to First Token (ms),534.83,309.14,2402.46,647678.44,321.15,343.79,357.53,384.77,433.10,631.45,791.98,982.34,1574.95,247.34
+Time to Second Token (ms),35.31,0.00,926.38,42765.60,0.00,0.00,0.16,16.54,23.37,28.40,48.75,67.77,334.56,76.22
+Tokens In Flight (tokens),86803.88,0.00,433946.31,,,,,,51804.04,,228793.33,276259.62,355469.65,93609.19
+Usage Completion Tokens (tokens),353.41,32.00,2265.00,427981.00,42.00,52.00,61.00,88.00,152.00,432.50,1105.00,1226.50,1599.10,411.11
+Usage Prompt Cache Read Tokens (tokens),26497.97,5824.00,85120.00,31877056.00,7744.00,8832.00,10816.00,16224.00,23424.00,34848.00,46105.60,50924.80,65914.88,13659.02
+Usage Prompt Tokens (tokens),27173.74,8094.00,85205.00,32907398.00,8445.70,9849.50,11809.00,16810.50,24222.00,35490.00,46635.00,51372.00,73467.30,13653.18
+Usage Total Tokens (tokens),27527.15,8203.00,85403.00,33335379.00,8521.10,9944.50,11920.00,17032.00,24621.00,35917.00,46899.00,51934.50,74108.80,13772.09
+
+Metric,Value
+Benchmark Duration (sec),2762.52
+Completed Requests (Success + Error),1213.00
+Error Request Count,2.00
+Input Token Throughput (tokens/sec),11912.05
+OSL Mismatch Count,1067.00
+Output Token Throughput (tokens/sec),154.92
+Overall Usage Prompt Cache Read % (%),96.87
+Replay Schedule Degraded,0.00
+Replay Schedule Lag p50 (ms),1.50
+Replay Schedule Lag p90 (ms),4.26
+Replay Schedule Lag p99 (ms),5.23
+Request Count,1211.00
+Request Error Rate (%),0.16
+Request Throughput (requests/sec),0.44
+Total Input Sequence Length (tokens),32907398.00
+Total Output Sequence Length (tokens),427981.00
+Total Output Tokens (tokens),427981.00
+Total Token Throughput (tokens/sec),12066.98
+Total Usage Completion Tokens (tokens),427981.00
+Total Usage Prompt Cache Read Tokens (tokens),31877056.00
+Total Usage Prompt Tokens (tokens),32907398.00
+Total Usage Total Tokens (tokens),33335379.00
+
+
+Endpoint,GPU_Index,GPU_Name,GPU_UUID,Metric,avg,min,max,sum,p1,p5,p10,p25,p50,p75,p90,p95,p99,std

experiments/kv-cache-fp8/capacity/paced/N0032/profile_export_aiperf.json

@@ -0,0 +1,987 @@
+{
+  "schema_version": "1.4",
+  "aiperf_version": "0.11.0",
+  "benchmark_id": "7966faaad416",
+  "request_throughput": {
+    "unit": "requests/sec",
+    "avg": 0.4383663875751429
+  },
+  "request_latency": {
+    "unit": "ms",
+    "avg": 6357.721891473162,
+    "p1": 890.5716237,
+    "p5": 1101.0463325,
+    "p10": 1267.252411,
+    "p25": 1878.8765179999998,
+    "p50": 3135.820107,
+    "p75": 7611.412791,
+    "p90": 17796.355036,
+    "p95": 21797.982572,
+    "p99": 27560.113839500034,
+    "min": 681.070244,
+    "max": 43595.130499,
+    "std": 6981.827347546926,
+    "count": 1211,
+    "sum": 7699201.210573999
+  },
+  "request_count": {
+    "unit": "requests",
+    "avg": 1211.0
+  },
+  "time_to_first_token": {
+    "unit": "ms",
+    "avg": 534.8294338769612,
+    "p1": 321.1527175,
+    "p5": 343.78526949999997,
+    "p10": 357.534118,
+    "p25": 384.7654215,
+    "p50": 433.10146599999996,
+    "p75": 631.4544404999999,
+    "p90": 791.9807179999999,
+    "p95": 982.335296,
+    "p99": 1574.948941500003,
+    "min": 309.14171999999996,
+    "max": 2402.4633679999997,
+    "std": 247.33999617012142,
+    "count": 1211,
+    "sum": 647678.444425
+  },
+  "time_to_second_token": {
+    "unit": "ms",
+    "avg": 35.31428773327828,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.156811,
+    "p25": 16.537633,
+    "p50": 23.367079,
+    "p75": 28.395121999999997,
+    "p90": 48.748304,
+    "p95": 67.7729785,
+    "p99": 334.5635116000001,
+    "min": 0.0,
+    "max": 926.3806989999999,
+    "std": 76.22254087081973,
+    "count": 1211,
+    "sum": 42765.602445
+  },
+  "inter_token_latency": {
+    "unit": "ms",
+    "avg": 16.440004270139532,
+    "p1": 9.520590782582122,
+    "p5": 10.159811625332697,
+    "p10": 10.749499048780487,
+    "p25": 12.74634114363323,
+    "p50": 15.87416808219178,
+    "p75": 19.174958144355415,
+    "p90": 22.131038226533168,
+    "p95": 24.387971376503124,
+    "p99": 31.298117415366185,
+    "min": 5.534955677966102,
+    "max": 39.5635622,
+    "std": 4.728546753946111,
+    "count": 1211,
+    "sum": 19908.84517113897
+  },
+  "output_token_throughput": {
+    "unit": "tokens/sec",
+    "avg": 154.92360439372192
+  },
+  "output_token_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 65.72263381015745,
+    "p1": 31.951481429297203,
+    "p5": 41.00400700003729,
+    "p10": 45.185408373706025,
+    "p25": 52.151474394415146,
+    "p50": 62.995427213715615,
+    "p75": 78.45388844670669,
+    "p90": 93.02759091024322,
+    "p95": 98.42707251044615,
+    "p99": 105.03595005568936,
+    "min": 25.27578267459445,
+    "max": 180.66992008280437,
+    "std": 18.203682726395307,
+    "count": 1211,
+    "sum": 79590.10954410068
+  },
+  "output_sequence_length": {
+    "unit": "tokens",
+    "avg": 353.411230388109,
+    "p1": 42.0,
+    "p5": 52.0,
+    "p10": 61.0,
+    "p25": 88.0,
+    "p50": 152.0,
+    "p75": 432.5,
+    "p90": 1105.0,
+    "p95": 1226.5,
+    "p99": 1599.1000000000026,
+    "min": 32.0,
+    "max": 2265.0,
+    "std": 411.1091251788751,
+    "count": 1211,
+    "sum": 427981.0
+  },
+  "input_sequence_length": {
+    "unit": "tokens",
+    "avg": 27173.73905862923,
+    "p1": 8445.7,
+    "p5": 9849.5,
+    "p10": 11809.0,
+    "p25": 16810.5,
+    "p50": 24222.0,
+    "p75": 35490.0,
+    "p90": 46635.0,
+    "p95": 51372.0,
+    "p99": 73467.30000000076,
+    "min": 8094.0,
+    "max": 85205.0,
+    "std": 13653.183483652007,
+    "count": 1211,
+    "sum": 32907398.0
+  },
+  "output_token_count": {
+    "unit": "tokens",
+    "avg": 353.411230388109,
+    "p1": 42.0,
+    "p5": 52.0,
+    "p10": 61.0,
+    "p25": 88.0,
+    "p50": 152.0,
+    "p75": 432.5,
+    "p90": 1105.0,
+    "p95": 1226.5,
+    "p99": 1599.1000000000026,
+    "min": 32.0,
+    "max": 2265.0,
+    "std": 411.1091251788751,
+    "count": 1211,
+    "sum": 427981.0
+  },
+  "inter_chunk_latency": {
+    "unit": "ms",
+    "avg": 30.90170893874018,
+    "p1": 0.0,
+    "p5": 0.09642400000000002,
+    "p10": 0.8073593,
+    "p25": 19.1645,
+    "p50": 24.6592745,
+    "p75": 27.817345,
+    "p90": 47.012340599999995,
+    "p95": 62.56494639999997,
+    "p99": 302.72282734,
+    "min": 0.0,
+    "max": 1693.717759,
+    "std": 52.20263106063583,
+    "count": 228192,
+    "sum": 7051522.766148999
+  },
+  "total_output_tokens": {
+    "unit": "tokens",
+    "avg": 427981.0
+  },
+  "benchmark_duration": {
+    "unit": "sec",
+    "avg": 2762.51917184
+  },
+  "total_isl": {
+    "unit": "tokens",
+    "avg": 32907398.0
+  },
+  "total_osl": {
+    "unit": "tokens",
+    "avg": 427981.0
+  },
+  "error_request_count": {
+    "unit": "requests",
+    "avg": 2.0
+  },
+  "telemetry_data": {
+    "summary": {
+      "endpoints_configured": [],
+      "endpoints_successful": [],
+      "start_time": "2026-07-20T17:08:09.219001",
+      "end_time": "2026-07-20T17:54:14.542157"
+    },
+    "endpoints": {},
+    "error_summary": []
+  },
+  "input_config": {
+    "models": {
+      "items": [
+        {
+          "name": "glm-5.2"
+        }
+      ]
+    },
+    "endpoint": {
+      "urls": [
+        "http://127.0.0.1:8391"
+      ],
+      "type": "chat",
+      "api_key": "<redacted>",
+      "streaming": true,
+      "use_server_token_count": true,
+      "extra": {
+        "cache_salt": "d513fc35125fa06a944a846295b1a179"
+      }
+    },
+    "datasets": [
+      {
+        "name": "main",
+        "type": "file",
+        "path": "results/capacity/sweep-paced-jul20/N0032/paced.jsonl",
+        "format": "dag_jsonl",
+        "entries": 1000000,
+        "random_seed": 0
+      }
+    ],
+    "phases": [
+      {
+        "name": "profiling",
+        "type": "fixed_schedule",
+        "requests": 1000000
+      }
+    ],
+    "artifacts": {
+      "dir": "/home/ubuntu/workspace/projects/glm-5.2-experiments/results/capacity/sweep-paced-jul20/N0032",
+      "records": [
+        "jsonl"
+      ],
+      "raw": true,
+      "trace": true,
+      "slice_duration": 60.0,
+      "auto_plot": false,
+      "plot_required": false
+    },
+    "tokenizer": {
+      "name": "builtin"
+    },
+    "gpu_telemetry": {
+      "enabled": true
+    },
+    "server_metrics": {
+      "enabled": true,
+      "urls": [],
+      "formats": [
+        "json",
+        "csv",
+        "jsonl"
+      ]
+    },
+    "network_latency": {
+      "enabled": false
+    },
+    "otel": {},
+    "mlflow": {},
+    "wandb": {},
+    "runtime": {
+      "ui": "simple"
+    }
+  },
+  "run_info": {
+    "benchmark_id": "7966faaad416",
+    "random_seed": 0,
+    "trial": 0,
+    "cli_command": "aiperf profile --model 'glm-5.2' --url 'http://127.0.0.1:8391' --api-key '<redacted>' --endpoint-type 'chat' --streaming --tokenizer 'builtin' --custom-dataset-type 'dag_jsonl' --input-file 'results/capacity/sweep-paced-jul20/N0032/paced.jsonl' --extra-inputs '{\"cache_salt\": \"d513fc35125fa06a944a846295b1a179\"}' --fixed-schedule --request-count 1000000 --random-seed 0 --use-server-token-count --artifact-dir 'results/capacity/sweep-paced-jul20/N0032' --export-level 'raw' --export-http-trace --slice-duration 60 --server-metrics-formats 'json' 'csv' 'jsonl' --ui 'simple'"
+  },
+  "was_cancelled": false,
+  "error_summary": [
+    {
+      "error_details": {
+        "type": "ServerDisconnectedError",
+        "message": "ServerDisconnectedError('Server disconnected')",
+        "cause_chain": [
+          "ServerDisconnectedError"
+        ]
+      },
+      "count": 2
+    }
+  ],
+  "start_time": "2026-07-20T17:08:09.219001",
+  "end_time": "2026-07-20T17:54:11.748324",
+  "http_req_chunks_received": {
+    "unit": "count",
+    "avg": 182.25516102394715,
+    "p1": 21.0,
+    "p5": 26.0,
+    "p10": 31.0,
+    "p25": 45.0,
+    "p50": 80.0,
+    "p75": 221.5,
+    "p90": 576.0,
+    "p95": 621.5,
+    "p99": 795.1000000000008,
+    "min": 15.0,
+    "max": 1207.0,
+    "std": 210.69549051550743,
+    "count": 1211,
+    "sum": 220711.0
+  },
+  "http_req_receiving": {
+    "unit": "ms",
+    "avg": 5859.049879396367,
+    "p1": 522.8906989,
+    "p5": 667.460457,
+    "p10": 831.286986,
+    "p25": 1389.2690535,
+    "p50": 2619.275139,
+    "p75": 7137.4053655,
+    "p90": 17348.565199,
+    "p95": 21276.2338145,
+    "p99": 26847.978621600018,
+    "min": 326.597685,
+    "max": 43209.377391999995,
+    "std": 6958.852036215867,
+    "count": 1211,
+    "sum": 7095309.403949
+  },
+  "http_req_dns_lookup": {
+    "unit": "ms",
+    "avg": 0.0,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 0.0,
+    "p75": 0.0,
+    "p90": 0.0,
+    "p95": 0.0,
+    "p99": 0.0,
+    "min": 0.0,
+    "max": 0.0,
+    "std": 0.0,
+    "count": 1211,
+    "sum": 0.0
+  },
+  "usage_total_tokens": {
+    "unit": "tokens",
+    "avg": 27527.15028901734,
+    "p1": 8521.1,
+    "p5": 9944.5,
+    "p10": 11920.0,
+    "p25": 17032.0,
+    "p50": 24621.0,
+    "p75": 35917.0,
+    "p90": 46899.0,
+    "p95": 51934.5,
+    "p99": 74108.80000000079,
+    "min": 8203.0,
+    "max": 85403.0,
+    "std": 13772.088601879082,
+    "count": 1211,
+    "sum": 33335379.0
+  },
+  "usage_prompt_tokens": {
+    "unit": "tokens",
+    "avg": 27173.73905862923,
+    "p1": 8445.7,
+    "p5": 9849.5,
+    "p10": 11809.0,
+    "p25": 16810.5,
+    "p50": 24222.0,
+    "p75": 35490.0,
+    "p90": 46635.0,
+    "p95": 51372.0,
+    "p99": 73467.30000000076,
+    "min": 8094.0,
+    "max": 85205.0,
+    "std": 13653.183483652007,
+    "count": 1211,
+    "sum": 32907398.0
+  },
+  "time_to_first_output_token": {
+    "unit": "ms",
+    "avg": 2533.5477450445633,
+    "p1": 356.13517406,
+    "p5": 378.13524735,
+    "p10": 400.04058879999997,
+    "p25": 450.2308235,
+    "p50": 802.6435315,
+    "p75": 2180.85330975,
+    "p90": 7177.949862599997,
+    "p95": 12000.73522025,
+    "p99": 20166.158505339994,
+    "min": 336.16898399999997,
+    "max": 28296.256973,
+    "std": 4123.086574450745,
+    "count": 1122,
+    "sum": 2842640.56994
+  },
+  "http_req_data_received": {
+    "unit": "KB",
+    "avg": 46.90052448905863,
+    "p1": 6.06064453125,
+    "p5": 7.3369140625,
+    "p10": 8.595703125,
+    "p25": 12.22265625,
+    "p50": 20.966796875,
+    "p75": 56.638671875,
+    "p90": 143.0302734375,
+    "p95": 159.79931640625,
+    "p99": 206.47255859375016,
+    "min": 4.5751953125,
+    "max": 292.4365234375,
+    "std": 53.23789090220943,
+    "count": 1211,
+    "sum": 56796.53515625
+  },
+  "http_req_connection_reused": {
+    "unit": "ratio",
+    "avg": 0.5920726672171759,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 1.0,
+    "p75": 1.0,
+    "p90": 1.0,
+    "p95": 1.0,
+    "p99": 1.0,
+    "min": 0.0,
+    "max": 1.0,
+    "std": 0.4914495131257281,
+    "count": 1211,
+    "sum": 717.0
+  },
+  "http_req_blocked": {
+    "unit": "ms",
+    "avg": 0.0,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 0.0,
+    "p75": 0.0,
+    "p90": 0.0,
+    "p95": 0.0,
+    "p99": 0.0,
+    "min": 0.0,
+    "max": 0.0,
+    "std": 0.0,
+    "count": 1211,
+    "sum": 0.0
+  },
+  "http_req_connecting": {
+    "unit": "ms",
+    "avg": 0.3884865185796862,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 0.0,
+    "p75": 0.7834685,
+    "p90": 0.9615859999999999,
+    "p95": 1.1743565,
+    "p99": 2.284598100000006,
+    "min": 0.0,
+    "max": 5.144407999999999,
+    "std": 0.584022600248336,
+    "count": 1211,
+    "sum": 470.45717399999995
+  },
+  "http_req_sending": {
+    "unit": "ms",
+    "avg": 0.6897690520231213,
+    "p1": 0.1862681,
+    "p5": 0.1990255,
+    "p10": 0.209662,
+    "p25": 0.2335265,
+    "p50": 0.30222899999999997,
+    "p75": 1.0669475,
+    "p90": 1.3330039999999999,
+    "p95": 1.718829,
+    "p99": 3.027531600000007,
+    "min": 0.16977199999999998,
+    "max": 6.145589999999999,
+    "std": 0.7001503127773558,
+    "count": 1211,
+    "sum": 835.3103219999999
+  },
+  "http_req_waiting": {
+    "unit": "ms",
+    "avg": 513.050246611891,
+    "p1": 318.67441560000003,
+    "p5": 336.5986565,
+    "p10": 349.691935,
+    "p25": 376.52982349999996,
+    "p50": 422.378061,
+    "p75": 620.951339,
+    "p90": 732.5862169999999,
+    "p95": 901.5415734999999,
+    "p99": 1494.8200766000025,
+    "min": 307.900775,
+    "max": 2402.105056,
+    "std": 219.7839924093609,
+    "count": 1211,
+    "sum": 621303.8486469999
+  },
+  "http_req_chunks_sent": {
+    "unit": "count",
+    "avg": 1.0,
+    "p1": 1.0,
+    "p5": 1.0,
+    "p10": 1.0,
+    "p25": 1.0,
+    "p50": 1.0,
+    "p75": 1.0,
+    "p90": 1.0,
+    "p95": 1.0,
+    "p99": 1.0,
+    "min": 1.0,
+    "max": 1.0,
+    "std": 0.0,
+    "count": 1211,
+    "sum": 1211.0
+  },
+  "http_req_duration": {
+    "unit": "ms",
+    "avg": 6372.78989506028,
+    "p1": 912.2532613,
+    "p5": 1104.2461205,
+    "p10": 1276.028036,
+    "p25": 1901.2105835,
+    "p50": 3151.3092979999997,
+    "p75": 7611.292918,
+    "p90": 17796.017304,
+    "p95": 21797.8571815,
+    "p99": 27559.96652840003,
+    "min": 680.9179899999999,
+    "max": 43594.977362,
+    "std": 6978.152547809003,
+    "count": 1211,
+    "sum": 7717448.562918
+  },
+  "usage_completion_tokens": {
+    "unit": "tokens",
+    "avg": 353.411230388109,
+    "p1": 42.0,
+    "p5": 52.0,
+    "p10": 61.0,
+    "p25": 88.0,
+    "p50": 152.0,
+    "p75": 432.5,
+    "p90": 1105.0,
+    "p95": 1226.5,
+    "p99": 1599.1000000000026,
+    "min": 32.0,
+    "max": 2265.0,
+    "std": 411.1091251788751,
+    "count": 1211,
+    "sum": 427981.0
+  },
+  "http_req_data_sent": {
+    "unit": "KB",
+    "avg": 111.22965911436829,
+    "p1": 38.21640625,
+    "p5": 43.41015625,
+    "p10": 49.083984375,
+    "p25": 69.81689453125,
+    "p50": 98.94921875,
+    "p75": 145.875,
+    "p90": 188.970703125,
+    "p95": 207.2353515625,
+    "p99": 290.3538085937529,
+    "min": 36.796875,
+    "max": 342.5947265625,
+    "std": 54.3807671706847,
+    "count": 1211,
+    "sum": 134699.1171875
+  },
+  "prefill_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 56624.96346519564,
+    "p1": 9392.37935087017,
+    "p5": 16690.601089123036,
+    "p10": 22060.362394744392,
+    "p25": 32265.967854883504,
+    "p50": 51225.154918337124,
+    "p75": 75475.5308287541,
+    "p90": 103016.314935781,
+    "p95": 114198.84604466811,
+    "p99": 130804.33602519821,
+    "min": 4963.687361825587,
+    "max": 174877.33418678862,
+    "std": 30523.405801688554,
+    "count": 1211,
+    "sum": 68572830.75635192
+  },
+  "e2e_output_token_throughput": {
+    "unit": "tokens/sec/user",
+    "avg": 52.94501843255356,
+    "p1": 25.351478240173897,
+    "p5": 32.40219099410561,
+    "p10": 36.13155272887695,
+    "p25": 43.370206347693326,
+    "p50": 50.97462932741093,
+    "p75": 61.3197942813713,
+    "p90": 72.87507534523675,
+    "p95": 80.5617380859373,
+    "p99": 91.2183619971803,
+    "min": 15.911347606645794,
+    "max": 97.2541292793307,
+    "std": 14.347996143546174,
+    "count": 1211,
+    "sum": 64116.41732182236
+  },
+  "osl_mismatch_diff_pct": {
+    "unit": "%",
+    "avg": -71.29388193555668,
+    "p1": -96.32522426186657,
+    "p5": -95.52102376599635,
+    "p10": -94.64428457234213,
+    "p25": -92.20677199938896,
+    "p50": -86.83982683982684,
+    "p75": -64.76543425909723,
+    "p90": 0.0,
+    "p95": 0.0,
+    "p99": 0.0,
+    "min": -97.52358490566037,
+    "max": 0.0,
+    "std": 31.550760274501204,
+    "count": 1211,
+    "sum": -86336.89102395915
+  },
+  "http_req_connection_overhead": {
+    "unit": "ms",
+    "avg": 0.3884865185796862,
+    "p1": 0.0,
+    "p5": 0.0,
+    "p10": 0.0,
+    "p25": 0.0,
+    "p50": 0.0,
+    "p75": 0.7834685,
+    "p90": 0.9615859999999999,
+    "p95": 1.1743565,
+    "p99": 2.284598100000006,
+    "min": 0.0,
+    "max": 5.144407999999999,
+    "std": 0.584022600248336,
+    "count": 1211,
+    "sum": 470.45717399999995
+  },
+  "http_req_total": {
+    "unit": "ms",
+    "avg": 6373.17838157886,
+    "p1": 912.5686896,
+    "p5": 1104.59374,
+    "p10": 1277.094769,
+    "p25": 1901.2105835,
+    "p50": 3151.3092979999997,
+    "p75": 7611.656465999999,
+    "p90": 17798.176044,
+    "p95": 21797.8571815,
+    "p99": 27559.96652840003,
+    "min": 681.8663349999999,
+    "max": 43595.898359,
+    "std": 6978.117628116199,
+    "count": 1211,
+    "sum": 7717919.020091999
+  },
+  "osl_mismatch_count": {
+    "unit": "requests",
+    "avg": 1067.0
+  },
+  "usage_prompt_cache_read_tokens": {
+    "unit": "tokens",
+    "avg": 26497.968412302576,
+    "p1": 7744.0,
+    "p5": 8832.0,
+    "p10": 10816.0,
+    "p25": 16224.0,
+    "p50": 23424.0,
+    "p75": 34848.0,
+    "p90": 46105.59999999998,
+    "p95": 50924.799999999974,
+    "p99": 65914.88,
+    "min": 5824.0,
+    "max": 85120.0,
+    "std": 13659.01883227155,
+    "count": 1203,
+    "sum": 31877056.0
+  },
+  "total_usage_prompt_tokens": {
+    "unit": "tokens",
+    "avg": 32907398.0
+  },
+  "total_usage_total_tokens": {
+    "unit": "tokens",
+    "avg": 33335379.0
+  },
+  "request_error_rate": {
+    "unit": "%",
+    "avg": 0.16488046166529266
+  },
+  "completed_request_count": {
+    "unit": "requests",
+    "avg": 1213.0
+  },
+  "total_usage_prompt_cache_read_tokens": {
+    "unit": "tokens",
+    "avg": 31877056.0
+  },
+  "total_usage_completion_tokens": {
+    "unit": "tokens",
+    "avg": 427981.0
+  },
+  "overall_usage_prompt_cache_read_pct": {
+    "unit": "%",
+    "avg": 96.86896545269244
+  },
+  "input_token_throughput": {
+    "unit": "tokens/sec",
+    "avg": 11912.053828040862
+  },
+  "total_token_throughput": {
+    "unit": "tokens/sec",
+    "avg": 12066.977432434584
+  },
+  "adj_time_to_first_token": {
+    "unit": "ms",
+    "p1": 321.12688499999996,
+    "p5": 343.869562,
+    "p10": 357.534118,
+    "p25": 384.90225999999996,
+    "p50": 433.640973,
+    "p75": 633.738086,
+    "p90": 794.0348409999999,
+    "p95": 984.0134109999999,
+    "p99": 1592.1480789999998,
+    "min": 309.14171999999996,
+    "count": 1213
+  },
+  "adj_request_latency": {
+    "unit": "ms",
+    "p1": 890.4087139999999,
+    "p5": 1102.268785,
+    "p10": 1267.252411,
+    "p25": 1879.596206,
+    "p50": 3146.361473,
+    "p75": 7763.189826999999,
+    "p90": 17973.891981,
+    "p95": 22040.741307,
+    "p99": 28301.960066,
+    "min": 681.070244,
+    "count": 1213
+  },
+  "adj_inter_token_latency": {
+    "unit": "ms",
+    "p1": 9.51399338961039,
+    "p5": 10.167106474999999,
+    "p10": 10.749499048780487,
+    "p25": 12.748403778463649,
+    "p50": 15.880178220248666,
+    "p75": 19.213954347826085,
+    "p90": 22.2638332375,
+    "p95": 24.472112918604648,
+    "p99": 31.467754853107344,
+    "min": 5.534955677966102,
+    "count": 1213
+  },
+  "effective_concurrency": {
+    "unit": "requests",
+    "avg": 2.7936812253089225,
+    "p50": 2.0,
+    "p90": 7.0,
+    "p95": 9.0,
+    "p99": 11.0,
+    "min": 0.0,
+    "max": 15.0,
+    "std": 3.012654357349507
+  },
+  "effective_decode_throughput": {
+    "unit": "tokens/sec",
+    "avg": 154.48520805554728,
+    "p50": 58.81871040721154,
+    "p90": 420.7976712308012,
+    "p95": 514.0390337433349,
+    "p99": 718.6839195143799,
+    "min": 0.0,
+    "max": 139868.71689933908,
+    "std": 531.6080201987124
+  },
+  "effective_prefill_throughput": {
+    "unit": "tokens/sec",
+    "avg": 11912.05382794598,
+    "p50": 0.0,
+    "p90": 58509.71715641835,
+    "p95": 85334.27409943833,
+    "p99": 126391.95046942095,
+    "min": 0.0,
+    "max": 222958.9263766172,
+    "std": 29899.915261746744
+  },
+  "effective_decode_concurrency": {
+    "unit": "requests",
+    "avg": 2.5592286974107763,
+    "p50": 2.0,
+    "p90": 7.0,
+    "p95": 8.0,
+    "p99": 11.0,
+    "min": 0.0,
+    "max": 15.0,
+    "std": 2.822833496494052
+  },
+  "effective_prefill_concurrency": {
+    "unit": "requests",
+    "avg": 0.23445124675337423,
+    "p50": 0.0,
+    "p90": 1.0,
+    "p95": 1.0,
+    "p99": 3.0,
+    "min": 0.0,
+    "max": 5.0,
+    "std": 0.5827643909967308
+  },
+  "effective_total_throughput": {
+    "unit": "tokens/sec",
+    "avg": 12066.539065950901,
+    "p50": 162.31116732456235,
+    "p90": 58677.50741063309,
+    "p95": 85553.13511870272,
+    "p99": 126709.14009161908,
+    "min": 0.0,
+    "max": 223087.15492198293,
+    "std": 29925.38865441848
+  },
+  "effective_decode_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 47.020774253721775,
+    "p50": 32.911752011744206,
+    "p90": 109.00109709523912,
+    "p95": 115.44372868734212,
+    "p99": 142.38922335087307,
+    "min": 0.0,
+    "max": 48841.47286831324,
+    "std": 146.87781108349446
+  },
+  "effective_prefill_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 9728.47445397068,
+    "p50": 0.0,
+    "p90": 43885.256571997685,
+    "p95": 70265.5166904017,
+    "p99": 112919.28634721976,
+    "min": 0.0,
+    "max": 174877.2778358074,
+    "std": 24971.716263977014
+  },
+  "tokens_in_flight": {
+    "unit": "tokens",
+    "avg": 86803.88499631052,
+    "p50": 51804.042801838645,
+    "p90": 228793.32503260954,
+    "p95": 276259.622615748,
+    "p99": 355469.64959154435,
+    "min": 0.0,
+    "max": 433946.31160636665,
+    "std": 93609.18601380874
+  },
+  "active_decode_throughput": {
+    "unit": "tokens/sec",
+    "avg": 227.28973717800238,
+    "p50": 129.42804556763673,
+    "p90": 482.1016654811991,
+    "p95": 562.7703105958831,
+    "p99": 851.2617916043152,
+    "min": 0.0,
+    "max": 139868.71689933908,
+    "std": 631.8576806410176
+  },
+  "active_prefill_throughput": {
+    "unit": "tokens/sec",
+    "avg": 68854.34533542546,
+    "p50": 65449.1206850116,
+    "p90": 115309.74688219285,
+    "p95": 127760.09680755845,
+    "p99": 170940.12831356996,
+    "min": 4963.687071896314,
+    "max": 222958.9263766172,
+    "std": 35310.40934012648
+  },
+  "active_decode_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 69.18034135019894,
+    "p50": 68.01200768694615,
+    "p90": 114.0185167022794,
+    "p95": 117.23238664529103,
+    "p99": 191.04996420543307,
+    "min": 0.0,
+    "max": 48841.47286831324,
+    "std": 173.80125310730625
+  },
+  "active_prefill_throughput_per_user": {
+    "unit": "tokens/sec/user",
+    "avg": 56232.76634874517,
+    "p50": 50516.26066312677,
+    "p90": 103409.90196536331,
+    "p95": 115483.01308115397,
+    "p99": 133523.6980545435,
+    "min": 4963.687071896314,
+    "max": 174877.2778358074,
+    "std": 31454.818971955337
+  },
+  "active_total_throughput": {
+    "unit": "tokens/sec",
+    "avg": 17206.866109929877,
+    "p50": 261.8212142773177,
+    "p90": 73526.17242836714,
+    "p95": 99462.72326149093,
+    "p99": 133927.04840339758,
+    "min": 36.567617549822984,
+    "max": 223087.15492198293,
+    "std": 34475.70002366313
+  },
+  "credit_to_start_latency": {
+    "unit": "ms",
+    "avg": 1.565462159934048,
+    "p1": 0.95228928,
+    "p5": 1.0493952,
+    "p10": 1.107968,
+    "p25": 1.219072,
+    "p50": 1.388544,
+    "p75": 1.613568,
+    "p90": 2.0248576,
+    "p95": 2.9318143999999964,
+    "p99": 4.724060159999983,
+    "min": 0.648704,
+    "max": 8.060416,
+    "std": 0.704949821889786,
+    "count": 1213,
+    "sum": 1898.9056
+  },
+  "effective_latency": {
+    "unit": "ms",
+    "avg": 6363.994402611706,
+    "p1": 909.90961664,
+    "p5": 1104.4664320000002,
+    "p10": 1276.4088319999998,
+    "p25": 1898.880256,
+    "p50": 3152.60928,
+    "p75": 7598.24384,
+    "p90": 17795.2436736,
+    "p95": 21776.73518079997,
+    "p99": 27554.607093759958,
+    "min": 3.29728,
+    "max": 43596.521216,
+    "std": 6977.178646967816,
+    "count": 1213,
+    "sum": 7719525.210368
+  },
+  "replay_sched_lag_p50": {
+    "unit": "ms",
+    "avg": 1.4983680000000001
+  },
+  "replay_sched_lag_p90": {
+    "unit": "ms",
+    "avg": 4.260121600000001
+  },
+  "replay_sched_lag_p99": {
+    "unit": "ms",
+    "avg": 5.229685760000001
+  },
+  "replay_sched_degraded": {
+    "unit": "count",
+    "avg": 0.0
+  }
+}

experiments/kv-cache-fp8/capacity/paced/N0032/profile_export_console.txt

@@ -0,0 +1,104 @@
+
+
+                               NVIDIA AIPerf | Error Summary                               
+┏━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━┓
+┃ Code ┃                    Type ┃ Message                                        ┃ Count ┃
+┡━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━┩
+│  N/A │ ServerDisconnectedError │ ServerDisconnectedError('Server disconnected') │     2 │
+└──────┴─────────────────────────┴────────────────────────────────────────────────┴───────┘
+
+
+                                            NVIDIA AIPerf | LLM Metrics: Effective                                             
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                    Metric ┃       avg ┃  min ┃        max ┃        p99 ┃        p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│          Effective Concurrency (requests) │      2.79 │ 0.00 │      15.00 │      11.00 │       7.00 │      2.00 │      3.01 │
+│  Effective Decode Throughput (tokens/sec) │    154.49 │ 0.00 │ 139,868.72 │     718.68 │     420.80 │     58.82 │    531.61 │
+│ Effective Prefill Throughput (tokens/sec) │ 11,912.05 │ 0.00 │ 222,958.93 │ 126,391.95 │  58,509.72 │      0.00 │ 29,899.92 │
+│   Effective Decode Concurrency (requests) │      2.56 │ 0.00 │      15.00 │      11.00 │       7.00 │      2.00 │      2.82 │
+│  Effective Prefill Concurrency (requests) │      0.23 │ 0.00 │       5.00 │       3.00 │       1.00 │      0.00 │      0.58 │
+│   Effective Total Throughput (tokens/sec) │ 12,066.54 │ 0.00 │ 223,087.15 │ 126,709.14 │  58,677.51 │    162.31 │ 29,925.39 │
+│      Effective Decode Throughput Per User │     47.02 │ 0.00 │  48,841.47 │     142.39 │     109.00 │     32.91 │    146.88 │
+│                         (tokens/sec/user) │           │      │            │            │            │           │           │
+│     Effective Prefill Throughput Per User │  9,728.47 │ 0.00 │ 174,877.28 │ 112,919.29 │  43,885.26 │      0.00 │ 24,971.72 │
+│                         (tokens/sec/user) │           │      │            │            │            │           │           │
+│                 Tokens In Flight (tokens) │ 86,803.88 │ 0.00 │ 433,946.31 │ 355,469.65 │ 228,793.33 │ 51,804.04 │ 93,609.19 │
+│         Effective Latency (CO-aware) (ms) │  6,363.99 │ 3.30 │  43,596.52 │  27,554.61 │  17,795.24 │  3,152.61 │  6,977.18 │
+└───────────────────────────────────────────┴───────────┴──────┴────────────┴────────────┴────────────┴───────────┴───────────┘
+                                              NVIDIA AIPerf | LLM Metrics: Active                                               
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                 Metric ┃       avg ┃      min ┃        max ┃        p99 ┃        p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━��━━╇━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│  Active Decode Throughput (tokens/sec) │    227.29 │     0.00 │ 139,868.72 │     851.26 │     482.10 │    129.43 │    631.86 │
+│ Active Prefill Throughput (tokens/sec) │ 68,854.35 │ 4,963.69 │ 222,958.93 │ 170,940.13 │ 115,309.75 │ 65,449.12 │ 35,310.41 │
+│      Active Decode Throughput Per User │     69.18 │     0.00 │  48,841.47 │     191.05 │     114.02 │     68.01 │    173.80 │
+│                      (tokens/sec/user) │           │          │            │            │            │           │           │
+│     Active Prefill Throughput Per User │ 56,232.77 │ 4,963.69 │ 174,877.28 │ 133,523.70 │ 103,409.90 │ 50,516.26 │ 31,454.82 │
+│                      (tokens/sec/user) │           │          │            │            │            │           │           │
+│   Active Total Throughput (tokens/sec) │ 17,206.87 │    36.57 │ 223,087.15 │ 133,927.05 │  73,526.17 │    261.82 │ 34,475.70 │
+└────────────────────────────────────────┴───────────┴──────────┴────────────┴────────────┴────────────┴───────────┴───────────┘
+                                                NVIDIA AIPerf | LLM Metrics: Usage                                                
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                  Metric ┃           avg ┃      min ┃       max ┃       p99 ┃       p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│            Usage Prompt Tokens (tokens) │     27,173.74 │ 8,094.00 │ 85,205.00 │ 73,467.30 │ 46,635.00 │ 24,222.00 │ 13,653.18 │
+│ Usage Prompt Cache Read Tokens (tokens) │     26,497.97 │ 5,824.00 │ 85,120.00 │ 65,914.88 │ 46,105.60 │ 23,424.00 │ 13,659.02 │
+│        Usage Completion Tokens (tokens) │        353.41 │    32.00 │  2,265.00 │  1,599.10 │  1,105.00 │    152.00 │    411.11 │
+│             Usage Total Tokens (tokens) │     27,527.15 │ 8,203.00 │ 85,403.00 │ 74,108.80 │ 46,899.00 │ 24,621.00 │ 13,772.09 │
+│      Total Usage Prompt Tokens (tokens) │ 32,907,398.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│    Total Usage Prompt Cache Read Tokens │ 31,877,056.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                (tokens) │               │          │           │           │           │           │           │
+│       Overall Usage Prompt Cache Read % │         96.87 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                     (%) │               │          │           │           │           │           │           │
+│  Total Usage Completion Tokens (tokens) │    427,981.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│       Total Usage Total Tokens (tokens) │ 33,335,379.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+└─────────────────────────────────────────┴───────────────┴──────────┴───────────┴───────────┴───────────┴───────────┴───────────┘
+                                                    NVIDIA AIPerf | LLM Metrics                                                     
+┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━┓
+┃                                        Metric ┃       avg ┃      min ┃       max ┃       p99 ┃       p90 ┃       p50 ┃       std ┃
+┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━┩
+│                      Time to First Token (ms) │    534.83 │   309.14 │  2,402.46 │  1,574.95 │    791.98 │    433.10 │    247.34 │
+│                     Time to Second Token (ms) │     35.31 │     0.00 │    926.38 │    334.56 │     48.75 │     23.37 │     76.22 │
+│               Time to First Output Token (ms) │  2,533.55 │   336.17 │ 28,296.26 │ 20,166.16 │  7,177.95 │    802.64 │  4,123.09 │
+│                          Request Latency (ms) │  6,357.72 │   681.07 │ 43,595.13 │ 27,560.11 │ 17,796.36 │  3,135.82 │  6,981.83 │
+│                      Inter Token Latency (ms) │     16.44 │     5.53 │     39.56 │     31.30 │     22.13 │     15.87 │      4.73 │
+│              Output Token Throughput Per User │     65.72 │    25.28 │    180.67 │    105.04 │     93.03 │     63.00 │     18.20 │
+│                             (tokens/sec/user) │           │          │           │           │           │           │           │
+│ E2E Output Token Throughput (tokens/sec/user) │     52.95 │    15.91 │     97.25 │     91.22 │     72.88 │     50.97 │     14.35 │
+│               Output Sequence Length (tokens) │    353.41 │    32.00 │  2,265.00 │  1,599.10 │  1,105.00 │    152.00 │    411.11 │
+│                Input Sequence Length (tokens) │ 27,173.74 │ 8,094.00 │ 85,205.00 │ 73,467.30 │ 46,635.00 │ 24,222.00 │ 13,653.18 │
+│          Output Token Throughput (tokens/sec) │    154.92 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│           Input Token Throughput (tokens/sec) │ 11,912.05 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│             Request Throughput (requests/sec) │      0.44 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│          Completed Requests (Success + Error) │  1,213.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                                    (requests) │           │          │           │           │           │           │           │
+│                        Request Error Rate (%) │      0.16 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│                      Request Count (requests) │  1,211.00 │      N/A │       N/A │       N/A │       N/A │       N/A │       N/A │
+│          Time to First Token (error-adjusted) │       N/A │   309.14 │       N/A │  1,592.15 │    794.03 │    433.64 │       N/A │
+│                                          (ms) │           │          │           │           │           │           │           │
+│              Request Latency (error-adjusted) │       N/A │   681.07 │       N/A │ 28,301.96 │ 17,973.89 │  3,146.36 │       N/A │
+│                                          (ms) │           │          │           │           │           │           │           │
+│          Inter Token Latency (error-adjusted) │       N/A │     5.53 │       N/A │     31.47 │     22.26 │     15.88 │       N/A │
+│                                          (ms) │           │          │           │           │           │           │           │
+└───────────────────────────────────────────────┴───────────┴──────────┴───────────┴───────────┴───────────┴───────────┴───────────┘
+
+
+No GPU telemetry data collected during the benchmarking run.
+
+╭───────────────────────────────── Output Sequence Length Mismatch Warning ─────────────────────────────────╮
+│  1,067 of 1,211 requests (88.1%) have output length differing from requested by more than the threshold.  │
+│  Threshold (tokens): min(requested x 5%, 50)                                                              │
+│  Average mismatch: -71.3%                                                                                 │
+│                                                                                                           │
+│  Why: Server hit EOS token before reaching requested output length.                                       │
+│                                                                                                           │
+│  Fix Options:                                                                                             │
+│    - --extra-inputs ignore_eos:true - Generate until max_tokens (vLLM, TensorRT-LLM)                      │
+│    - --extra-inputs min_tokens:<N> - Set minimum output length (vLLM, TensorRT-LLM, SGLang)               │
+│    - --use-server-token-count - Use server-reported token counts if tokenizer mismatch suspected          │
+│                                                                                                           │
+│  Diagnostics:                                                                                             │
+│    - Review profile_export.jsonl -> osl_mismatch_diff_pct for per-request values                          │
+│    - Adjust: AIPERF_METRICS_OSL_MISMATCH_PCT_THRESHOLD=5                                                  │
+│    - Adjust: AIPERF_METRICS_OSL_MISMATCH_MAX_TOKEN_THRESHOLD=50                                           │
+╰───────────────────────────────────────────────────────────────────────────────────────────────────────────╯

experiments/kv-cache-fp8/capacity/paced/N0032/summary.json

@@ -0,0 +1,25 @@
+{
+  "sent": 1213,
+  "records": 1213,
+  "valid": 1211,
+  "valid_frac": 0.9984,
+  "p25_speed_tok_s": 52.1,
+  "p50_speed_tok_s": 63.0,
+  "p95_ttft_ms": 984.0,
+  "p50_ttft_ms": 433.0,
+  "cache_read_ratio": 0.9687,
+  "max_turn_index": 49,
+  "conversations": 32,
+  "recycled_record_frac": 0.0,
+  "inflight_mean": 3.7,
+  "inflight_p95": 11.0,
+  "foreign_p50": 0.0,
+  "foreign_p95": 3.0,
+  "active_sessions_mean": 11.2,
+  "active_sessions_p95": 28.0,
+  "tiers": {
+    "20tok": true,
+    "60tok": false,
+    "180tok": false
+  }
+}

experiments/kv-cache-fp8/capacity/paced/N0032/verify_cache_salt.log

@@ -0,0 +1 @@
+records_checked=1213 mismatches=0