Gahow Wang 1cf03c6e79 Cost model: add interference penalty for co-located heavy prefill ...

Old cost model: offload_cost = colocated_cost + RDMA_overhead, so offload
was always 0.1s more expensive. Result: only 19/117 HEAVY offloaded.

New: colocated_cost includes interference penalty when C_s has decode
requests: penalty = prefill_time × min(num_requests, 3) × 0.3.
Offload now wins when C_s has 1+ concurrent request.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-05-23 23:59:06 +08:00

analysis

Add comprehensive research findings document

2026-05-23 07:16:31 +08:00

experiments

Add elastic PS evaluation plan for production-realistic trace

2026-05-23 15:56:05 +08:00

patches

Add vLLM patches directory for version-controlled patch management

2026-05-22 00:26:14 +08:00

replayer

replayer: wire --max-inflight-sessions cap into replay loop (B2)

2026-05-23 21:04:09 +08:00

scripts

Cost model: add interference penalty for co-located heavy prefill

2026-05-23 23:59:06 +08:00

tests

proxy: Settings dataclass + cache-ratio gate + P-pick offload penalty (B4, M2, M3, D5)

2026-05-23 21:11:17 +08:00

third_party/vllm

Fix scheduler assertion crash on partial remote prefill finished_recving

2026-05-23 23:33:26 +08:00

.gitignore

Phase 1 milestone: system-level analysis + reproducible report

2026-05-22 16:17:41 +08:00

FIXES.md

Add FIXES.md with prioritized repo cleanup checklist

2026-05-23 20:35:56 +08:00

pyproject.toml

tests: add minimal coverage for percentile + proxy routing (S1)

2026-05-23 21:07:14 +08:00

REPORT.md

Report §3.8: Direct RDMA read results — HEAVY TTFT -70%, TPOT p90 -38%

2026-05-23 22:56:16 +08:00

TODO.md

LMetric routing policy (OSDI'26) + A/B results vs linear baseline

2026-05-22 16:57:32 +08:00

Description

No description provided

48 MiB

Languages

Python 82.9%

Shell 17.1%