Gahow Wang
d5b276180d
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>
23 lines
732 B
JSON
23 lines
732 B
JSON
{
|
|
"accounting_mode": "graceful-footer",
|
|
"cell": "tp1_mns16",
|
|
"invariants": {
|
|
"anchor_intervals_present": true,
|
|
"compile_capture_pre_ready": true,
|
|
"encoded_balanced": true,
|
|
"footer_sidecar_agrees": true,
|
|
"last_step_matches": true,
|
|
"layer1_contiguous": true,
|
|
"layer1_zero_drops": true,
|
|
"one_footer_last": true,
|
|
"one_ready_marker": true,
|
|
"sidecar_final": true,
|
|
"warmup_exact_16": true,
|
|
"warmup_long": true,
|
|
"written_matches_records": true
|
|
},
|
|
"layer1_records": 14174,
|
|
"post_ready_capture_events": [],
|
|
"stream": "/home/admin/cpfs/wjh/opprof-phase6-dash0-20260712/runs/phase6/solo-authoritative/cells/tp1_mns16/opprof/opprof-v1-dp0-pid2781473-1783876972093434233.jsonl"
|
|
}
|