Workload-conditioned operator profiling on patched vLLM 0.24.0 + Qwen3-30B-A3B/H20. H1b PASS (irregular patterns carry +23-45pp R64 raggedness, 8-45% token-efficiency loss vs rectangular controls); mechanism decomposition kills the padding narrative and finds the arrival-uniformization artifact (-12.9%); cross-version churn surface shows TP2/MNS64 -29.4% across vLLM 0.20->0.24 while the argmax held. Raw Layer-1 JSONL streams (507 MB) stay on disk, git-ignored; footer sidecars and metrics are tracked. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
18 lines
273 B
JSON
18 lines
273 B
JSON
{
|
|
"invariants": {
|
|
"counts_match": true,
|
|
"observations_92": true
|
|
},
|
|
"mismatches": [],
|
|
"observations": 92,
|
|
"request_counts": {
|
|
"distinct_n": 34,
|
|
"finite_n": 92,
|
|
"max": 600.0,
|
|
"min": 66.0,
|
|
"missing_n": 0,
|
|
"n": 92
|
|
},
|
|
"schema": 1
|
|
}
|