2.2 KiB
EXP-SIMFID-Q30-FIXED-PRESSURE-SURFACE: high-pressure Fixed-PD / Fixed-PO
Status: approved and launching (2026-07-19).
Question
When uniform Fixed-PD is moved from the near-singleton workload to
4096 -> 256 at 1.125 request/s/GPU, does Frontier still choose a different
latency-optimal configuration from real community vLLM 0.20? Does removing
decode (4096 -> 1) change that selection boundary?
Setup and decision rule
- Qwen3-30B-A3B BF16 on dash0 H20; community vLLM 0.20.0.
- Fixed-PD
4096 -> 256and Fixed-PO4096 -> 1; 257 uniform, prefix-disjoint requests; prefix caching off. - Rate is frozen at 1.125 request/s/GPU, so global rates are 1.125, 2.25, and 4.5 request/s for TP1, TP2, and TP4.
- Surface:
TP in {1,2,4}xMNS in {8,16,32,64}, MBT=8192. Real measurements use three fresh-server trials with rotated order. - Frontier is
deadc4a,piecewise, the frozen CUDA-event profile plus the graph-compatible KERNEL_ONLY profile and measured collectives used by the prior Trace-PD comparison. No latency calibration is fitted to this case. - Compare exact winners, selected-config real regret, and pairwise order for mean/p90 TTFT, E2E, and TPOT where decode exists. SLO is not scored.
The pressure-probe measurements selected the workload but are excluded from the real evaluation pool. A failed or incomplete Frontier cell is a coverage failure, not a high-latency observation.
Expected output and decision
The final table has one row per case/objective with Frontier winner, real winner, regret, and pairwise agreement. If the old Fixed-PD reversal persists, the failure is not an artifact of singleton load. If it disappears, simulator fidelity has a load-regime boundary that must be localized. Fixed-PO isolates whether decode-state composition is necessary for either outcome.
The previously reviewed matrix schematic remains the figure prototype:
../simulator-tuning-latency-matrix-v0/latency-selection-matrix-schematic.svg.
Provenance
Remote output root:
dash0:/home/admin/cpfs/wjh/aituner/qwen30-fixed-pressure-surface-20260719-r1.
The campaign records source/profile/model hashes, runtime versions, GPU
inventory, per-cell commands, raw request records, and an artifact checksum
manifest.