gahow/aituner
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Evaluate vLLM 0.20 profiles against Frontier

Browse Source
This commit is contained in:
Gahow Wang
2026-07-16 23:59:10 +08:00
parent 008324e70c
commit 76107d3e87
33 changed files with 13973 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

469
runs/frontier-qwen30-vllm020-profile-v1/artifacts/trace-support/support.json Normal file
View File

@@ -0,0 +1,469 @@
{
"block_size": 64,
"cohort_output": "/home/admin/cpfs/wjh/frontier_qwen30_vllm020_profiles/trace-support/cohort.jsonl",
"cohort_sha256": "9ef15a4c8843127a5c52ff8455bdeb9945ac6921f2977dd11070ed6944eed7c8",
"full_downsampled_cohort": {
"hash_blocks_per_request": {
"max": 128.0,
"min": 2.0,
"p10": 4.0,
"p25": 15.0,
"p50": 38.0,
"p75": 82.0,
"p90": 111.0,
"p95": 122.0,
"p99": 127.0
},
"input_tokens": {
"max": 8186.0,
"min": 71.0,
"p10": 208.0,
"p25": 933.0,
"p50": 2421.0,
"p75": 5244.0,
"p90": 7103.0,
"p95": 7775.0,
"p99": 8074.0
},
"interarrival_s": {
"max": 0.20259999999998257,
"min": 0.06279999999997088,
"p10": 0.0917999999999708,
"p25": 0.10240000000003846,
"p50": 0.1163999999999703,
"p75": 0.13009999999999167,
"p90": 0.14210000000002765,
"p95": 0.15119999999997447,
"p99": 0.17959999999997933
},
"multi_turn_fraction": 0.140625,
"output_tokens": {
"max": 128.0,
"min": 128.0,
"p10": 128.0,
"p25": 128.0,
"p50": 128.0,
"p75": 128.0,
"p90": 128.0,
"p95": 128.0,
"p99": 128.0
},
"prefix_reuse_upper_bound": {
"any_position_reusable_block_ratio": 0.09274807669777795,
"leading_reusable_blocks_per_request": {
"max": 55.0,
"min": 0.0,
"p10": 0.0,
"p25": 0.0,
"p50": 1.0,
"p75": 13.0,
"p90": 13.0,
"p95": 13.0,
"p99": 49.0
},
"leading_reusable_tokens_per_request": {
"max": 3520.0,
"min": 0.0,
"p10": 0.0,
"p25": 0.0,
"p50": 64.0,
"p75": 832.0,
"p90": 832.0,
"p95": 832.0,
"p99": 3136.0
},
"rows_with_hash_ids": 512,
"semantics": "arrival-ordered, infinite-capacity/no-eviction upper bound; not an observed KV-cache hit rate",
"total_blocks": 25607,
"unique_blocks": 23232
},
"request_count": 512,
"sampling_u": {
"max": 0.9979913161468553,
"min": 0.0017297900650431776,
"p10": 0.10365756051317851,
"p25": 0.22877823992809823,
"p50": 0.4667483893139273,
"p75": 0.7338360613813898,
"p90": 0.8969698510725516,
"p95": 0.9558957048665733,
"p99": 0.991130680217395
}
},
"limits": [
"Trace length/hash support does not determine dynamic decode or mixed batch shapes.",
"Those shapes depend jointly on arrival history, SLO pressure, TP execution time, MNS, chunking, and KV eviction.",
"MoE expert routing is not present in the trace and must be measured from model execution."
],
"loader_contract": {
"completion_tokens_override": 128,
"downsampling": "AITuner _downsample_requests before sampling_u threshold",
"input_length_filter": {
"max": 8192,
"min": 0
},
"max_requests_per_probe": 512,
"ordering": "arrival_s",
"replay_time_scale": 0.1
},
"schema_version": "qwen30_trace_profile_support.v1",
"study": "/home/admin/cpfs/wjh/aituner/aituner-qwen30-vllm020-profile-v1/configs/examples/dash0_qwen30b_a3b_community_vllm020_noharness.json",
"study_sha256": "cd0b9e2385751478f459faeae2cf3d12c58229866176911051421bc043bf99d4",
"threshold_cohorts": {
"0.125": {
"hash_blocks_per_request": {
"max": 127.0,
"min": 2.0,
"p10": 7.0,
"p25": 15.0,
"p50": 30.0,
"p75": 76.0,
"p90": 105.0,
"p95": 111.0,
"p99": 127.0
},
"input_tokens": {
"max": 8110.0,
"min": 74.0,
"p10": 392.0,
"p25": 922.0,
"p50": 1864.0,
"p75": 4844.0,
"p90": 6686.0,
"p95": 7054.0,
"p99": 8110.0
},
"interarrival_s": {
"max": 6.217200000000002,
"min": 0.10639999999998517,
"p10": 0.16680000000001272,
"p25": 0.30810000000001736,
"p50": 0.6001999999999938,
"p75": 1.1201000000000016,
"p90": 1.8216999999999643,
"p95": 2.419699999999967,
"p99": 6.217200000000002
},
"multi_turn_fraction": 0.15151515151515152,
"output_tokens": {
"max": 128.0,
"min": 128.0,
"p10": 128.0,
"p25": 128.0,
"p50": 128.0,
"p75": 128.0,
"p90": 128.0,
"p95": 128.0,
"p99": 128.0
},
"prefix_reuse_upper_bound": {
"any_position_reusable_block_ratio": 0.10466666666666667,
"leading_reusable_blocks_per_request": {
"max": 13.0,
"min": 0.0,
"p10": 0.0,
"p25": 0.0,
"p50": 1.0,
"p75": 13.0,
"p90": 13.0,
"p95": 13.0,
"p99": 13.0
},
"leading_reusable_tokens_per_request": {
"max": 832.0,
"min": 0.0,
"p10": 0.0,
"p25": 0.0,
"p50": 64.0,
"p75": 832.0,
"p90": 832.0,
"p95": 832.0,
"p99": 832.0
},
"rows_with_hash_ids": 66,
"semantics": "arrival-ordered, infinite-capacity/no-eviction upper bound; not an observed KV-cache hit rate",
"total_blocks": 3000,
"unique_blocks": 2686
},
"request_count": 66,
"sampling_u": {
"max": 0.12321850600323363,
"min": 0.0017297900650431776,
"p10": 0.01599560440427848,
"p25": 0.03636425746001375,
"p50": 0.06451196858246099,
"p75": 0.10162658487005978,
"p90": 0.1138971596689021,
"p95": 0.12001874701158237,
"p99": 0.12321850600323363
}
},
"0.25": {
"hash_blocks_per_request": {
"max": 128.0,
"min": 2.0,
"p10": 4.0,
"p25": 14.0,
"p50": 34.0,
"p75": 87.0,
"p90": 107.0,
"p95": 122.0,
"p99": 127.0
},
"input_tokens": {
"max": 8149.0,
"min": 72.0,
"p10": 197.0,
"p25": 867.0,
"p50": 2167.0,
"p75": 5568.0,
"p90": 6841.0,
"p95": 7745.0,
"p99": 8110.0
},
"interarrival_s": {
"max": 2.3110999999999855,
"min": 0.07340000000003499,
"p10": 0.11439999999997497,
"p25": 0.14210000000002765,
"p50": 0.3411000000000044,
"p75": 0.5650000000000093,
"p90": 0.8215000000000146,
"p95": 1.088799999999992,
"p99": 1.7577999999999996
},
"multi_turn_fraction": 0.11888111888111888,
"output_tokens": {
"max": 128.0,
"min": 128.0,
"p10": 128.0,
"p25": 128.0,
"p50": 128.0,
"p75": 128.0,
"p90": 128.0,
"p95": 128.0,
"p99": 128.0
},
"prefix_reuse_upper_bound": {
"any_position_reusable_block_ratio": 0.07953890489913544,
"leading_reusable_blocks_per_request": {
"max": 13.0,
"min": 0.0,
"p10": 0.0,
"p25": 0.0,
"p50": 1.0,
"p75": 13.0,
"p90": 13.0,
"p95": 13.0,
"p99": 13.0
},
"leading_reusable_tokens_per_request": {
"max": 832.0,
"min": 0.0,
"p10": 0.0,
"p25": 0.0,
"p50": 64.0,
"p75": 832.0,
"p90": 832.0,
"p95": 832.0,
"p99": 832.0
},
"rows_with_hash_ids": 143,
"semantics": "arrival-ordered, infinite-capacity/no-eviction upper bound; not an observed KV-cache hit rate",
"total_blocks": 6940,
"unique_blocks": 6388
},
"request_count": 143,
"sampling_u": {
"max": 0.24941730181025415,
"min": 0.0017297900650431776,
"p10": 0.03095529459918134,
"p25": 0.07431293234041883,
"p50": 0.13301201741611407,
"p75": 0.1994148366998705,
"p90": 0.2292686392189136,
"p95": 0.24107176388894497,
"p99": 0.24671643798343437
}
},
"0.5": {
"hash_blocks_per_request": {
"max": 128.0,
"min": 2.0,
"p10": 2.0,
"p25": 15.0,
"p50": 42.0,
"p75": 86.0,
"p90": 109.0,
"p95": 117.0,
"p99": 127.0
},
"input_tokens": {
"max": 8149.0,
"min": 71.0,
"p10": 113.0,
"p25": 905.0,
"p50": 2648.0,
"p75": 5490.0,
"p90": 6926.0,
"p95": 7476.0,
"p99": 8074.0
},
"interarrival_s": {
"max": 1.2561000000000142,
"min": 0.06749999999997236,
"p10": 0.10019999999999563,
"p25": 0.11840000000002249,
"p50": 0.15819999999994394,
"p75": 0.2599999999999483,
"p90": 0.3867999999999938,
"p95": 0.5083999999999804,
"p99": 0.696700000000007
},
"multi_turn_fraction": 0.14130434782608695,
"output_tokens": {
"max": 128.0,
"min": 128.0,
"p10": 128.0,
"p25": 128.0,
"p50": 128.0,
"p75": 128.0,
"p90": 128.0,
"p95": 128.0,
"p99": 128.0
},
"prefix_reuse_upper_bound": {
"any_position_reusable_block_ratio": 0.08318264014466546,
"leading_reusable_blocks_per_request": {
"max": 52.0,
"min": 0.0,
"p10": 0.0,
"p25": 0.0,
"p50": 1.0,
"p75": 13.0,
"p90": 13.0,
"p95": 13.0,
"p99": 24.0
},
"leading_reusable_tokens_per_request": {
"max": 3328.0,
"min": 0.0,
"p10": 0.0,
"p25": 0.0,
"p50": 64.0,
"p75": 832.0,
"p90": 832.0,
"p95": 832.0,
"p99": 1536.0
},
"rows_with_hash_ids": 276,
"semantics": "arrival-ordered, infinite-capacity/no-eviction upper bound; not an observed KV-cache hit rate",
"total_blocks": 13825,
"unique_blocks": 12675
},
"request_count": 276,
"sampling_u": {
"max": 0.49832600818075473,
"min": 0.0017297900650431776,
"p10": 0.056494245471656455,
"p25": 0.1313132001019981,
"p50": 0.24403265916374703,
"p75": 0.3706937275751198,
"p90": 0.4530058956953501,
"p95": 0.48171254599865826,
"p99": 0.49639910750753047
}
},
"1.0": {
"hash_blocks_per_request": {
"max": 128.0,
"min": 2.0,
"p10": 4.0,
"p25": 15.0,
"p50": 38.0,
"p75": 82.0,
"p90": 111.0,
"p95": 122.0,
"p99": 127.0
},
"input_tokens": {
"max": 8186.0,
"min": 71.0,
"p10": 208.0,
"p25": 933.0,
"p50": 2421.0,
"p75": 5244.0,
"p90": 7103.0,
"p95": 7775.0,
"p99": 8074.0
},
"interarrival_s": {
"max": 0.20259999999998257,
"min": 0.06279999999997088,
"p10": 0.0917999999999708,
"p25": 0.10240000000003846,
"p50": 0.1163999999999703,
"p75": 0.13009999999999167,
"p90": 0.14210000000002765,
"p95": 0.15119999999997447,
"p99": 0.17959999999997933
},
"multi_turn_fraction": 0.140625,
"output_tokens": {
"max": 128.0,
"min": 128.0,
"p10": 128.0,
"p25": 128.0,
"p50": 128.0,
"p75": 128.0,
"p90": 128.0,
"p95": 128.0,
"p99": 128.0
},
"prefix_reuse_upper_bound": {
"any_position_reusable_block_ratio": 0.09274807669777795,
"leading_reusable_blocks_per_request": {
"max": 55.0,
"min": 0.0,
"p10": 0.0,
"p25": 0.0,
"p50": 1.0,
"p75": 13.0,
"p90": 13.0,
"p95": 13.0,
"p99": 49.0
},
"leading_reusable_tokens_per_request": {
"max": 3520.0,
"min": 0.0,
"p10": 0.0,
"p25": 0.0,
"p50": 64.0,
"p75": 832.0,
"p90": 832.0,
"p95": 832.0,
"p99": 3136.0
},
"rows_with_hash_ids": 512,
"semantics": "arrival-ordered, infinite-capacity/no-eviction upper bound; not an observed KV-cache hit rate",
"total_blocks": 25607,
"unique_blocks": 23232
},
"request_count": 512,
"sampling_u": {
"max": 0.9979913161468553,
"min": 0.0017297900650431776,
"p10": 0.10365756051317851,
"p25": 0.22877823992809823,
"p50": 0.4667483893139273,
"p75": 0.7338360613813898,
"p90": 0.8969698510725516,
"p95": 0.9558957048665733,
"p99": 0.991130680217395
}
}
},
"trace": "/home/admin/cpfs/wjh/aituner/aituner/trace_windows/traces/chat_w20260311_1000.jsonl",
"trace_sha256": "f539f38eb0ee0f750e3c23ff47df6eed3faf723a25f1444d55665a85871750b9",
"window_id": "chat_w20260311_1000"
}