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

87
runs/frontier-qwen30-vllm020-profile-v1/artifacts/qwen30-initial-op-trace-delta.json Normal file
View File

@@ -0,0 +1,87 @@
{
"comparison_contract": {
"calibration_a_tp": 1.0,
"fixture": "fidelity_p1_tp1_mns64_low1",
"scheduler_state": "batch_id=0, layer_id=0 before profile-dependent trajectories diverge"
},
"layer_component_sum_ms": {
"historical_profile": 1.148334,
"new_over_old": 2.700194368537377,
"vllm020_profile": 3.1007249999999997
},
"rows": [
{
"component": "input_layernorm",
"historical_profile_ms": 0.010619,
"new_over_old": 1.955833882663151,
"vllm020_profile_ms": 0.020769
},
{
"component": "attn_pre_proj",
"historical_profile_ms": 0.204389,
"new_over_old": 1.0199276869107439,
"vllm020_profile_ms": 0.208462
},
{
"component": "attn_rope",
"historical_profile_ms": 0.018908,
"new_over_old": 7.394171779141103,
"vllm020_profile_ms": 0.139809
},
{
"component": "attn_kv_cache_save",
"historical_profile_ms": 0.014226,
"new_over_old": 1.427386475467454,
"vllm020_profile_ms": 0.020306
},
{
"component": "attn_prefill",
"historical_profile_ms": 0.107759,
"new_over_old": 2.670709639102071,
"vllm020_profile_ms": 0.287793
},
{
"component": "attn_decode",
"historical_profile_ms": 0.0,
"new_over_old": null,
"vllm020_profile_ms": 0.0
},
{
"component": "attn_post_proj",
"historical_profile_ms": 0.134153,
"new_over_old": 0.967559428413826,
"vllm020_profile_ms": 0.129801
},
{
"component": "post_attention_layernorm",
"historical_profile_ms": 0.011481,
"new_over_old": 0.9000087100426792,
"vllm020_profile_ms": 0.010333
},
{
"component": "moe_gating_linear",
"historical_profile_ms": 0.011802,
"new_over_old": 2.2642772411455687,
"vllm020_profile_ms": 0.026723
},
{
"component": "moe_gating_routing_topk",
"historical_profile_ms": 0.009662,
"new_over_old": 2.2922790312564687,
"vllm020_profile_ms": 0.022148
},
{
"component": "moe_shuffling",
"historical_profile_ms": 0.030431,
"new_over_old": 0.0,
"vllm020_profile_ms": 0.0
},
{
"component": "moe_grouped_gemm",
"historical_profile_ms": 0.594904,
"new_over_old": 3.756204362384519,
"vllm020_profile_ms": 2.234581
}
],
"schema": "frontier-initial-op-trace-delta.v1"
}

37
runs/frontier-qwen30-vllm020-profile-v1/artifacts/qwen30-p1-profile-ablation.csv Normal file
View File

@@ -0,0 +1,37 @@
mode,cell,role,real_feasible,sim_feasible,real_pass_rate,sim_pass_rate,pass_rate_absolute_error,offered_req_s_per_gpu,sim_throughput_req_s_per_gpu
historical-calibrated,tp1_mns64,high1,False,True,0.24581005586592178,0.9888268156424581,0.7430167597765364,2.9833333333333334,2.8823345959582403
historical-calibrated,tp1_mns64,low1,True,True,1.0,1.0,0.0,2.033333333333333,1.9806287931485522
historical-calibrated,tp1_mns8,high1,False,False,0.20670391061452514,0.16759776536312848,0.03910614525139666,2.9833333333333334,2.162661799726284
historical-calibrated,tp1_mns8,low1,True,True,1.0,0.9754098360655737,0.024590163934426257,2.033333333333333,1.9806313066623409
historical-calibrated,tp2_mns64,high1,True,True,1.0,1.0,0.0,2.875,2.809583898291986
historical-calibrated,tp2_mns64,low1,True,True,1.0,1.0,0.0,1.9583333333333333,1.911740196570019
historical-calibrated,tp2_mns8,high1,False,False,0.20238095238095238,0.15476190476190477,0.047619047619047616,2.8,2.083217332903509
historical-calibrated,tp2_mns8,low1,True,True,1.0,0.9868995633187773,0.013100436681222738,1.9083333333333334,1.8637394408880168
historical-calibrated,tp4_mns16,high1,False,False,0.10666666666666667,0.2693333333333333,0.16266666666666663,3.125,2.6829294600763345
historical-calibrated,tp4_mns16,low1,False,True,0.6196078431372549,1.0,0.3803921568627451,2.125,2.0865667413849516
historical-calibrated,tp4_mns64,high1,True,True,1.0,1.0,0.0,3.125,3.0718260341342285
historical-calibrated,tp4_mns64,low1,True,True,1.0,1.0,0.0,2.125,2.0967062495193307
historical-profile-only,tp1_mns64,high1,False,False,0.24581005586592178,0.24581005586592178,0.0,2.9833333333333334,2.82546810606982
historical-profile-only,tp1_mns64,low1,True,False,1.0,0.7377049180327869,0.2622950819672131,2.033333333333333,1.9427723213024963
historical-profile-only,tp1_mns8,high1,False,False,0.20670391061452514,0.09497206703910614,0.111731843575419,2.9833333333333334,1.5728783582441022
historical-profile-only,tp1_mns8,low1,True,False,1.0,0.1721311475409836,0.8278688524590164,2.033333333333333,1.5629164685645112
historical-profile-only,tp2_mns64,high1,True,False,1.0,0.2608695652173913,0.7391304347826086,2.875,2.7121507155597584
historical-profile-only,tp2_mns64,low1,True,False,1.0,0.8382978723404255,0.16170212765957448,1.9583333333333333,1.849101894293262
historical-profile-only,tp2_mns8,high1,False,False,0.20238095238095238,0.041666666666666664,0.16071428571428573,2.8,0.9859710483293953
historical-profile-only,tp2_mns8,low1,True,False,1.0,0.0611353711790393,0.9388646288209607,1.9083333333333334,0.9771135206177947
historical-profile-only,tp4_mns16,high1,False,False,0.10666666666666667,0.028,0.07866666666666668,3.125,0.9604809617478992
historical-profile-only,tp4_mns16,low1,False,False,0.6196078431372549,0.058823529411764705,0.5607843137254902,2.125,0.9495221159963836
historical-profile-only,tp4_mns64,high1,True,False,1.0,0.0026666666666666666,0.9973333333333333,3.125,2.281094127221267
historical-profile-only,tp4_mns64,low1,True,False,1.0,0.4627450980392157,0.5372549019607843,2.125,1.9793960278798877
vllm020-profile-only,tp1_mns64,high1,False,False,0.24581005586592178,0.0,0.24581005586592178,2.9833333333333334,2.074240230678028
vllm020-profile-only,tp1_mns64,low1,True,False,1.0,0.0,1.0,2.033333333333333,1.6189525827713862
vllm020-profile-only,tp1_mns8,high1,False,False,0.20670391061452514,0.0,0.20670391061452514,2.9833333333333334,0.5271192714936628
vllm020-profile-only,tp1_mns8,low1,True,False,1.0,0.0,1.0,2.033333333333333,0.5225514701978402
vllm020-profile-only,tp2_mns64,high1,True,False,1.0,0.0,1.0,2.875,1.7237810220598582
vllm020-profile-only,tp2_mns64,low1,True,False,1.0,0.0,1.0,1.9583333333333333,1.608919184557297
vllm020-profile-only,tp2_mns8,high1,False,False,0.20238095238095238,0.0,0.20238095238095238,2.8,0.39315152956277477
vllm020-profile-only,tp2_mns8,low1,True,False,1.0,0.0,1.0,1.9083333333333334,0.3894201655965407
vllm020-profile-only,tp4_mns16,high1,False,False,0.10666666666666667,0.0,0.10666666666666667,3.125,0.43700447014112276
vllm020-profile-only,tp4_mns16,low1,False,False,0.6196078431372549,0.0,0.6196078431372549,2.125,0.43475717185302865
vllm020-profile-only,tp4_mns64,high1,True,False,1.0,0.0,1.0,3.125,1.2598449015844073
vllm020-profile-only,tp4_mns64,low1,True,False,1.0,0.0,1.0,2.125,1.2025831680037373
1 mode cell role real_feasible sim_feasible real_pass_rate sim_pass_rate pass_rate_absolute_error offered_req_s_per_gpu sim_throughput_req_s_per_gpu
2 historical-calibrated tp1_mns64 high1 False True 0.24581005586592178 0.9888268156424581 0.7430167597765364 2.9833333333333334 2.8823345959582403
3 historical-calibrated tp1_mns64 low1 True True 1.0 1.0 0.0 2.033333333333333 1.9806287931485522
4 historical-calibrated tp1_mns8 high1 False False 0.20670391061452514 0.16759776536312848 0.03910614525139666 2.9833333333333334 2.162661799726284
5 historical-calibrated tp1_mns8 low1 True True 1.0 0.9754098360655737 0.024590163934426257 2.033333333333333 1.9806313066623409
6 historical-calibrated tp2_mns64 high1 True True 1.0 1.0 0.0 2.875 2.809583898291986
7 historical-calibrated tp2_mns64 low1 True True 1.0 1.0 0.0 1.9583333333333333 1.911740196570019
8 historical-calibrated tp2_mns8 high1 False False 0.20238095238095238 0.15476190476190477 0.047619047619047616 2.8 2.083217332903509
9 historical-calibrated tp2_mns8 low1 True True 1.0 0.9868995633187773 0.013100436681222738 1.9083333333333334 1.8637394408880168
10 historical-calibrated tp4_mns16 high1 False False 0.10666666666666667 0.2693333333333333 0.16266666666666663 3.125 2.6829294600763345
11 historical-calibrated tp4_mns16 low1 False True 0.6196078431372549 1.0 0.3803921568627451 2.125 2.0865667413849516
12 historical-calibrated tp4_mns64 high1 True True 1.0 1.0 0.0 3.125 3.0718260341342285
13 historical-calibrated tp4_mns64 low1 True True 1.0 1.0 0.0 2.125 2.0967062495193307
14 historical-profile-only tp1_mns64 high1 False False 0.24581005586592178 0.24581005586592178 0.0 2.9833333333333334 2.82546810606982
15 historical-profile-only tp1_mns64 low1 True False 1.0 0.7377049180327869 0.2622950819672131 2.033333333333333 1.9427723213024963
16 historical-profile-only tp1_mns8 high1 False False 0.20670391061452514 0.09497206703910614 0.111731843575419 2.9833333333333334 1.5728783582441022
17 historical-profile-only tp1_mns8 low1 True False 1.0 0.1721311475409836 0.8278688524590164 2.033333333333333 1.5629164685645112
18 historical-profile-only tp2_mns64 high1 True False 1.0 0.2608695652173913 0.7391304347826086 2.875 2.7121507155597584
19 historical-profile-only tp2_mns64 low1 True False 1.0 0.8382978723404255 0.16170212765957448 1.9583333333333333 1.849101894293262
20 historical-profile-only tp2_mns8 high1 False False 0.20238095238095238 0.041666666666666664 0.16071428571428573 2.8 0.9859710483293953
21 historical-profile-only tp2_mns8 low1 True False 1.0 0.0611353711790393 0.9388646288209607 1.9083333333333334 0.9771135206177947
22 historical-profile-only tp4_mns16 high1 False False 0.10666666666666667 0.028 0.07866666666666668 3.125 0.9604809617478992
23 historical-profile-only tp4_mns16 low1 False False 0.6196078431372549 0.058823529411764705 0.5607843137254902 2.125 0.9495221159963836
24 historical-profile-only tp4_mns64 high1 True False 1.0 0.0026666666666666666 0.9973333333333333 3.125 2.281094127221267
25 historical-profile-only tp4_mns64 low1 True False 1.0 0.4627450980392157 0.5372549019607843 2.125 1.9793960278798877
26 vllm020-profile-only tp1_mns64 high1 False False 0.24581005586592178 0.0 0.24581005586592178 2.9833333333333334 2.074240230678028
27 vllm020-profile-only tp1_mns64 low1 True False 1.0 0.0 1.0 2.033333333333333 1.6189525827713862
28 vllm020-profile-only tp1_mns8 high1 False False 0.20670391061452514 0.0 0.20670391061452514 2.9833333333333334 0.5271192714936628
29 vllm020-profile-only tp1_mns8 low1 True False 1.0 0.0 1.0 2.033333333333333 0.5225514701978402
30 vllm020-profile-only tp2_mns64 high1 True False 1.0 0.0 1.0 2.875 1.7237810220598582
31 vllm020-profile-only tp2_mns64 low1 True False 1.0 0.0 1.0 1.9583333333333333 1.608919184557297
32 vllm020-profile-only tp2_mns8 high1 False False 0.20238095238095238 0.0 0.20238095238095238 2.8 0.39315152956277477
33 vllm020-profile-only tp2_mns8 low1 True False 1.0 0.0 1.0 1.9083333333333334 0.3894201655965407
34 vllm020-profile-only tp4_mns16 high1 False False 0.10666666666666667 0.0 0.10666666666666667 3.125 0.43700447014112276
35 vllm020-profile-only tp4_mns16 low1 False False 0.6196078431372549 0.0 0.6196078431372549 2.125 0.43475717185302865
36 vllm020-profile-only tp4_mns64 high1 True False 1.0 0.0 1.0 3.125 1.2598449015844073
37 vllm020-profile-only tp4_mns64 low1 True False 1.0 0.0 1.0 2.125 1.2025831680037373

511
runs/frontier-qwen30-vllm020-profile-v1/artifacts/qwen30-p1-profile-ablation.json Normal file
View File

@@ -0,0 +1,511 @@
{
"rows": [
{
"cell": "tp1_mns64",
"mode": "historical-calibrated",
"offered_req_s_per_gpu": 2.9833333333333334,
"pass_rate_absolute_error": 0.7430167597765364,
"real_feasible": false,
"real_pass_rate": 0.24581005586592178,
"role": "high1",
"sim_feasible": true,
"sim_pass_rate": 0.9888268156424581,
"sim_throughput_req_s_per_gpu": 2.8823345959582403
},
{
"cell": "tp1_mns64",
"mode": "historical-calibrated",
"offered_req_s_per_gpu": 2.033333333333333,
"pass_rate_absolute_error": 0.0,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": true,
"sim_pass_rate": 1.0,
"sim_throughput_req_s_per_gpu": 1.9806287931485522
},
{
"cell": "tp1_mns8",
"mode": "historical-calibrated",
"offered_req_s_per_gpu": 2.9833333333333334,
"pass_rate_absolute_error": 0.03910614525139666,
"real_feasible": false,
"real_pass_rate": 0.20670391061452514,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.16759776536312848,
"sim_throughput_req_s_per_gpu": 2.162661799726284
},
{
"cell": "tp1_mns8",
"mode": "historical-calibrated",
"offered_req_s_per_gpu": 2.033333333333333,
"pass_rate_absolute_error": 0.024590163934426257,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": true,
"sim_pass_rate": 0.9754098360655737,
"sim_throughput_req_s_per_gpu": 1.9806313066623409
},
{
"cell": "tp2_mns64",
"mode": "historical-calibrated",
"offered_req_s_per_gpu": 2.875,
"pass_rate_absolute_error": 0.0,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "high1",
"sim_feasible": true,
"sim_pass_rate": 1.0,
"sim_throughput_req_s_per_gpu": 2.809583898291986
},
{
"cell": "tp2_mns64",
"mode": "historical-calibrated",
"offered_req_s_per_gpu": 1.9583333333333333,
"pass_rate_absolute_error": 0.0,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": true,
"sim_pass_rate": 1.0,
"sim_throughput_req_s_per_gpu": 1.911740196570019
},
{
"cell": "tp2_mns8",
"mode": "historical-calibrated",
"offered_req_s_per_gpu": 2.8,
"pass_rate_absolute_error": 0.047619047619047616,
"real_feasible": false,
"real_pass_rate": 0.20238095238095238,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.15476190476190477,
"sim_throughput_req_s_per_gpu": 2.083217332903509
},
{
"cell": "tp2_mns8",
"mode": "historical-calibrated",
"offered_req_s_per_gpu": 1.9083333333333334,
"pass_rate_absolute_error": 0.013100436681222738,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": true,
"sim_pass_rate": 0.9868995633187773,
"sim_throughput_req_s_per_gpu": 1.8637394408880168
},
{
"cell": "tp4_mns16",
"mode": "historical-calibrated",
"offered_req_s_per_gpu": 3.125,
"pass_rate_absolute_error": 0.16266666666666663,
"real_feasible": false,
"real_pass_rate": 0.10666666666666667,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.2693333333333333,
"sim_throughput_req_s_per_gpu": 2.6829294600763345
},
{
"cell": "tp4_mns16",
"mode": "historical-calibrated",
"offered_req_s_per_gpu": 2.125,
"pass_rate_absolute_error": 0.3803921568627451,
"real_feasible": false,
"real_pass_rate": 0.6196078431372549,
"role": "low1",
"sim_feasible": true,
"sim_pass_rate": 1.0,
"sim_throughput_req_s_per_gpu": 2.0865667413849516
},
{
"cell": "tp4_mns64",
"mode": "historical-calibrated",
"offered_req_s_per_gpu": 3.125,
"pass_rate_absolute_error": 0.0,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "high1",
"sim_feasible": true,
"sim_pass_rate": 1.0,
"sim_throughput_req_s_per_gpu": 3.0718260341342285
},
{
"cell": "tp4_mns64",
"mode": "historical-calibrated",
"offered_req_s_per_gpu": 2.125,
"pass_rate_absolute_error": 0.0,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": true,
"sim_pass_rate": 1.0,
"sim_throughput_req_s_per_gpu": 2.0967062495193307
},
{
"cell": "tp1_mns64",
"mode": "historical-profile-only",
"offered_req_s_per_gpu": 2.9833333333333334,
"pass_rate_absolute_error": 0.0,
"real_feasible": false,
"real_pass_rate": 0.24581005586592178,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.24581005586592178,
"sim_throughput_req_s_per_gpu": 2.82546810606982
},
{
"cell": "tp1_mns64",
"mode": "historical-profile-only",
"offered_req_s_per_gpu": 2.033333333333333,
"pass_rate_absolute_error": 0.2622950819672131,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": false,
"sim_pass_rate": 0.7377049180327869,
"sim_throughput_req_s_per_gpu": 1.9427723213024963
},
{
"cell": "tp1_mns8",
"mode": "historical-profile-only",
"offered_req_s_per_gpu": 2.9833333333333334,
"pass_rate_absolute_error": 0.111731843575419,
"real_feasible": false,
"real_pass_rate": 0.20670391061452514,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.09497206703910614,
"sim_throughput_req_s_per_gpu": 1.5728783582441022
},
{
"cell": "tp1_mns8",
"mode": "historical-profile-only",
"offered_req_s_per_gpu": 2.033333333333333,
"pass_rate_absolute_error": 0.8278688524590164,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": false,
"sim_pass_rate": 0.1721311475409836,
"sim_throughput_req_s_per_gpu": 1.5629164685645112
},
{
"cell": "tp2_mns64",
"mode": "historical-profile-only",
"offered_req_s_per_gpu": 2.875,
"pass_rate_absolute_error": 0.7391304347826086,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.2608695652173913,
"sim_throughput_req_s_per_gpu": 2.7121507155597584
},
{
"cell": "tp2_mns64",
"mode": "historical-profile-only",
"offered_req_s_per_gpu": 1.9583333333333333,
"pass_rate_absolute_error": 0.16170212765957448,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": false,
"sim_pass_rate": 0.8382978723404255,
"sim_throughput_req_s_per_gpu": 1.849101894293262
},
{
"cell": "tp2_mns8",
"mode": "historical-profile-only",
"offered_req_s_per_gpu": 2.8,
"pass_rate_absolute_error": 0.16071428571428573,
"real_feasible": false,
"real_pass_rate": 0.20238095238095238,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.041666666666666664,
"sim_throughput_req_s_per_gpu": 0.9859710483293953
},
{
"cell": "tp2_mns8",
"mode": "historical-profile-only",
"offered_req_s_per_gpu": 1.9083333333333334,
"pass_rate_absolute_error": 0.9388646288209607,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": false,
"sim_pass_rate": 0.0611353711790393,
"sim_throughput_req_s_per_gpu": 0.9771135206177947
},
{
"cell": "tp4_mns16",
"mode": "historical-profile-only",
"offered_req_s_per_gpu": 3.125,
"pass_rate_absolute_error": 0.07866666666666668,
"real_feasible": false,
"real_pass_rate": 0.10666666666666667,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.028,
"sim_throughput_req_s_per_gpu": 0.9604809617478992
},
{
"cell": "tp4_mns16",
"mode": "historical-profile-only",
"offered_req_s_per_gpu": 2.125,
"pass_rate_absolute_error": 0.5607843137254902,
"real_feasible": false,
"real_pass_rate": 0.6196078431372549,
"role": "low1",
"sim_feasible": false,
"sim_pass_rate": 0.058823529411764705,
"sim_throughput_req_s_per_gpu": 0.9495221159963836
},
{
"cell": "tp4_mns64",
"mode": "historical-profile-only",
"offered_req_s_per_gpu": 3.125,
"pass_rate_absolute_error": 0.9973333333333333,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.0026666666666666666,
"sim_throughput_req_s_per_gpu": 2.281094127221267
},
{
"cell": "tp4_mns64",
"mode": "historical-profile-only",
"offered_req_s_per_gpu": 2.125,
"pass_rate_absolute_error": 0.5372549019607843,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": false,
"sim_pass_rate": 0.4627450980392157,
"sim_throughput_req_s_per_gpu": 1.9793960278798877
},
{
"cell": "tp1_mns64",
"mode": "vllm020-profile-only",
"offered_req_s_per_gpu": 2.9833333333333334,
"pass_rate_absolute_error": 0.24581005586592178,
"real_feasible": false,
"real_pass_rate": 0.24581005586592178,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.0,
"sim_throughput_req_s_per_gpu": 2.074240230678028
},
{
"cell": "tp1_mns64",
"mode": "vllm020-profile-only",
"offered_req_s_per_gpu": 2.033333333333333,
"pass_rate_absolute_error": 1.0,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": false,
"sim_pass_rate": 0.0,
"sim_throughput_req_s_per_gpu": 1.6189525827713862
},
{
"cell": "tp1_mns8",
"mode": "vllm020-profile-only",
"offered_req_s_per_gpu": 2.9833333333333334,
"pass_rate_absolute_error": 0.20670391061452514,
"real_feasible": false,
"real_pass_rate": 0.20670391061452514,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.0,
"sim_throughput_req_s_per_gpu": 0.5271192714936628
},
{
"cell": "tp1_mns8",
"mode": "vllm020-profile-only",
"offered_req_s_per_gpu": 2.033333333333333,
"pass_rate_absolute_error": 1.0,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": false,
"sim_pass_rate": 0.0,
"sim_throughput_req_s_per_gpu": 0.5225514701978402
},
{
"cell": "tp2_mns64",
"mode": "vllm020-profile-only",
"offered_req_s_per_gpu": 2.875,
"pass_rate_absolute_error": 1.0,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.0,
"sim_throughput_req_s_per_gpu": 1.7237810220598582
},
{
"cell": "tp2_mns64",
"mode": "vllm020-profile-only",
"offered_req_s_per_gpu": 1.9583333333333333,
"pass_rate_absolute_error": 1.0,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": false,
"sim_pass_rate": 0.0,
"sim_throughput_req_s_per_gpu": 1.608919184557297
},
{
"cell": "tp2_mns8",
"mode": "vllm020-profile-only",
"offered_req_s_per_gpu": 2.8,
"pass_rate_absolute_error": 0.20238095238095238,
"real_feasible": false,
"real_pass_rate": 0.20238095238095238,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.0,
"sim_throughput_req_s_per_gpu": 0.39315152956277477
},
{
"cell": "tp2_mns8",
"mode": "vllm020-profile-only",
"offered_req_s_per_gpu": 1.9083333333333334,
"pass_rate_absolute_error": 1.0,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": false,
"sim_pass_rate": 0.0,
"sim_throughput_req_s_per_gpu": 0.3894201655965407
},
{
"cell": "tp4_mns16",
"mode": "vllm020-profile-only",
"offered_req_s_per_gpu": 3.125,
"pass_rate_absolute_error": 0.10666666666666667,
"real_feasible": false,
"real_pass_rate": 0.10666666666666667,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.0,
"sim_throughput_req_s_per_gpu": 0.43700447014112276
},
{
"cell": "tp4_mns16",
"mode": "vllm020-profile-only",
"offered_req_s_per_gpu": 2.125,
"pass_rate_absolute_error": 0.6196078431372549,
"real_feasible": false,
"real_pass_rate": 0.6196078431372549,
"role": "low1",
"sim_feasible": false,
"sim_pass_rate": 0.0,
"sim_throughput_req_s_per_gpu": 0.43475717185302865
},
{
"cell": "tp4_mns64",
"mode": "vllm020-profile-only",
"offered_req_s_per_gpu": 3.125,
"pass_rate_absolute_error": 1.0,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "high1",
"sim_feasible": false,
"sim_pass_rate": 0.0,
"sim_throughput_req_s_per_gpu": 1.2598449015844073
},
{
"cell": "tp4_mns64",
"mode": "vllm020-profile-only",
"offered_req_s_per_gpu": 2.125,
"pass_rate_absolute_error": 1.0,
"real_feasible": true,
"real_pass_rate": 1.0,
"role": "low1",
"sim_feasible": false,
"sim_pass_rate": 0.0,
"sim_throughput_req_s_per_gpu": 1.2025831680037373
}
],
"schema": "frontier-qwen30-p1-profile-ablation.v1",
"scope": {
"cells": 6,
"probes_per_cell": 2,
"reading": "held-out boundary classification, not a complete capacity sweep",
"roles": [
"low1",
"high1"
]
},
"sources": {
"controller_state": "/home/gahow/phd/replayserve/runs/fidelity_p1_frontier_committed_20260714/real/p1b/controller-state.json",
"historical_calibrated": "/home/gahow/phd/replayserve/runs/fidelity_p1_frontier_committed_20260714/results/metrics.json",
"historical_profile_only": "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/comparison/old-profile-only-v2/results/metrics.json",
"vllm020_profile_only": "/home/gahow/phd/aituner/runs/frontier-qwen30-vllm020-profile-v1/comparison/new-profile-only-v2/results/metrics.json"
},
"summaries": {
"historical-calibrated": {
"feasible_probe_count": 9,
"p1_capacity_lower_bounds_req_s_per_gpu": {
"tp1_mns64": 2.9833333333333334,
"tp1_mns8": 2.033333333333333,
"tp2_mns64": 2.875,
"tp2_mns8": 1.9083333333333334,
"tp4_mns16": 2.125,
"tp4_mns64": 3.125
},
"pass_rate_mae": 0.11754094806600345,
"probe_classification": {
"accuracy": 0.8333333333333334,
"agreement": 10,
"false_feasible": 2,
"false_infeasible": 0
},
"rank_identifiable": true
},
"historical-profile-only": {
"feasible_probe_count": 0,
"p1_capacity_lower_bounds_req_s_per_gpu": {
"tp1_mns64": 0.0,
"tp1_mns8": 0.0,
"tp2_mns64": 0.0,
"tp2_mns8": 0.0,
"tp4_mns16": 0.0,
"tp4_mns64": 0.0
},
"pass_rate_mae": 0.44802887255544604,
"probe_classification": {
"accuracy": 0.4166666666666667,
"agreement": 5,
"false_feasible": 0,
"false_infeasible": 7
},
"rank_identifiable": false
},
"vllm020-profile-only": {
"feasible_probe_count": 0,
"p1_capacity_lower_bounds_req_s_per_gpu": {
"tp1_mns64": 0.0,
"tp1_mns8": 0.0,
"tp2_mns64": 0.0,
"tp2_mns8": 0.0,
"tp4_mns16": 0.0,
"tp4_mns64": 0.0
},
"pass_rate_mae": 0.69843078572211,
"probe_classification": {
"accuracy": 0.4166666666666667,
"agreement": 5,
"false_feasible": 0,
"false_infeasible": 7
},
"rank_identifiable": false
}
}
}

2171
runs/frontier-qwen30-vllm020-profile-v1/artifacts/qwen30-p1b-opprof-summary.json Normal file
View File

File diff suppressed because it is too large Load Diff

1853
runs/frontier-qwen30-vllm020-profile-v1/artifacts/qwen30-routing-mismatch.json Normal file
View File

File diff suppressed because it is too large Load Diff

5222
runs/frontier-qwen30-vllm020-profile-v1/artifacts/qwen30-s2-profile-ablation.json Normal file
View File

File diff suppressed because it is too large Load Diff

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"
}