gahow/agentic-kvc
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Commit Graph

  • ea5c3bfe6b compute_roofline: argparse --trace, fix stale default path (D4) Gahow Wang 2026-05-23 20:58:09 +08:00
  • 547611e022 scripts: archive obsolete one-off shell/python scripts to legacy/ (D2, D3) Gahow Wang 2026-05-23 20:57:32 +08:00
  • c64b0b39c7 bench.sh: fix stale MODEL and TRACE defaults (B6) Gahow Wang 2026-05-23 20:56:40 +08:00
  • 556f3011c6 proxy: remove dead state and broken fire-and-forget path (B1, D1) Gahow Wang 2026-05-23 20:56:11 +08:00
  • fc445df0ad Add FIXES.md with prioritized repo cleanup checklist Gahow Wang 2026-05-23 20:35:56 +08:00
  • b2ede1da77 bench.sh: add trap for graceful cleanup on kill/interrupt Gahow Wang 2026-05-23 20:24:13 +08:00
  • ea5149726c Partial remote prefill: C_s exports cache, D computes new tokens locally Gahow Wang 2026-05-23 20:04:13 +08:00
  • be273f7f27 Replace static offload gate with runtime cost model Gahow Wang 2026-05-23 19:42:33 +08:00
  • 9835d6af5d Elastic PS eval: near-neutral, offload gate triggers only 14% of HEAVY Gahow Wang 2026-05-23 16:49:25 +08:00
  • 03e88b30bd Add elastic PS evaluation plan for production-realistic trace Gahow Wang 2026-05-23 15:56:05 +08:00
  • f5e45afd4e Fix 4 elastic PS bugs: D accounting, offload cap, cache migration, prefix sync Gahow Wang 2026-05-23 15:55:11 +08:00
  • bf037594c4 Production-realistic baseline: APC 67.5%, TPOT +139% from interference Gahow Wang 2026-05-23 15:44:34 +08:00
  • d8dc9dc0ce Add --max-single-turn-ratio to control single-turn session fraction Gahow Wang 2026-05-23 14:17:25 +08:00
  • 1e1e2e774d Fix sampler: window+thin preserves cross-session KV cache sharing Gahow Wang 2026-05-23 14:03:12 +08:00
  • 4089ffd63f Fix replay methodology: trace-driven dispatch, no artificial limits Gahow Wang 2026-05-23 12:43:41 +08:00
  • c8ba666517 Benchmark concurrency gap: 1 req/GPU is 10-15x below production Gahow Wang 2026-05-23 12:16:20 +08:00
  • fefbd71ca9 GPU imbalance analysis + elastic PS verdict + corrected LMetric results Gahow Wang 2026-05-23 12:11:23 +08:00
  • 3594f7dce0 Fix LMetric routing: remove session affinity, align with OSDI'26 spec Gahow Wang 2026-05-23 11:56:58 +08:00
  • 8e0c6e78b0 Add comprehensive research findings document Gahow Wang 2026-05-23 07:16:31 +08:00
  • 080a8fa138 Chunk-size ablation + comprehensive synthesis Gahow Wang 2026-05-23 07:15:02 +08:00
  • baf7ffb08c 16-session contention: TPOT +45% from prefill-decode interference Gahow Wang 2026-05-23 05:51:47 +08:00
  • 85b230455e H7 OVERLOAD_FACTOR sweep: negative result + H4 GPU profiling Gahow Wang 2026-05-23 03:04:02 +08:00
  • 3bc37cc6d5 PS experiments + H4 cache-gate + GPU profiling + Mooncake elif→if fix Gahow Wang 2026-05-23 02:14:37 +08:00
  • 098d86385a Add elastic hypotheses tracking doc with H1-H6 analysis Gahow Wang 2026-05-23 01:17:12 +08:00
  • fc92410ec9 Invalidate prior A/B results + add proper experiment harness Gahow Wang 2026-05-22 17:54:21 +08:00
  • e4fa56cb1e LMetric routing policy (OSDI'26) + A/B results vs linear baseline Gahow Wang 2026-05-22 16:57:32 +08:00
  • 2b0ac70ee7 Phase 1 milestone: system-level analysis + reproducible report Gahow Wang 2026-05-22 16:17:41 +08:00
  • 1e8628581b Fair A/B: Elastic P2P wins on ALL metrics vs baseline (fresh restart) Gahow Wang 2026-05-22 15:48:51 +08:00
  • 76ee28a40f Elastic P2P v4: error rate 25% -> 4%, TTFT p50 -12% (median-tail tradeoff) Gahow Wang 2026-05-22 15:08:16 +08:00
  • 1d2eeb4925 Elastic P2P offload: TTFT p50 -49% vs baseline (0.551 vs 1.080) Gahow Wang 2026-05-22 13:50:25 +08:00
  • e9e313f9c5 P2P cache analysis: external KV correctly registered in prefix cache Gahow Wang 2026-05-22 13:25:34 +08:00
  • 1b9268ba4c P2P prefill offload: TTFT p50 -13% but p90 +59% (median-vs-tail tradeoff) Gahow Wang 2026-05-22 12:28:24 +08:00
  • 7f93d36970 System profile: 4 mechanisms why PD-Sep loses to session-sticky combined Gahow Wang 2026-05-22 10:58:59 +08:00
  • 42bcd31976 TP=2 DP=4 + hybrid routing: best TTFT at cost of TPOT Gahow Wang 2026-05-22 10:35:18 +08:00
  • a65ec42467 Update report: adaptive v2 confirms no KV transfer helps single-machine Gahow Wang 2026-05-22 10:15:08 +08:00
  • 2fee355626 Adaptive v2 (selective Mooncake offload): worse than baseline Gahow Wang 2026-05-22 10:14:10 +08:00
  • 4bf0b999ff Final GPU comparison: hybrid routing matches baseline latency with better APC Gahow Wang 2026-05-22 03:14:05 +08:00
  • 795edc6c66 Overnight work report: routing optimization achieves +4.7pp APC Gahow Wang 2026-05-22 02:54:48 +08:00
  • 012d73f596 Hybrid routing: session-sticky + load-aware override achieves best results Gahow Wang 2026-05-22 02:53:44 +08:00
  • efe984477a Balanced routing result: APC +4pp but latency +23% (cache-load tradeoff) Gahow Wang 2026-05-22 02:13:15 +08:00
  • 32f09d32cd Balanced session-sticky routing + agentic workload pattern analysis Gahow Wang 2026-05-22 01:50:27 +08:00
  • e45f00eb68 Cache policy simulation: routing quality dominates, not eviction policy Gahow Wang 2026-05-22 01:28:53 +08:00
  • 10636b1ab1 KV cache lifecycle design + eviction loss analysis Gahow Wang 2026-05-22 01:27:22 +08:00
  • d11d9f5cb9 Adaptive prefill offload v1: implementation + experiment Gahow Wang 2026-05-22 01:00:10 +08:00
  • d6e47d3742 Design doc: Adaptive Prefill Offload Gahow Wang 2026-05-22 00:44:22 +08:00
  • 445e491123 Add vLLM v0.18.1 source tree with KV transfer abort fix Gahow Wang 2026-05-22 00:30:38 +08:00
  • b6591950bc Add vLLM patches directory for version-controlled patch management Gahow Wang 2026-05-22 00:26:14 +08:00
  • efa70f05b5 Consolidate analysis into single report with appendix Gahow Wang 2026-05-22 00:23:23 +08:00
  • ce616f46d1 Add per-request breakdown profiling, identify KV cache memory bottleneck Gahow Wang 2026-05-22 00:13:50 +08:00
  • c7afdc5074 Ablation 2: fire-and-forget vs await-prefill scheduling Gahow Wang 2026-05-21 23:02:42 +08:00
  • 9dee25907b Add P/D ratio ablation: 6P+2D vs 4P+4D vs Combined Gahow Wang 2026-05-21 22:42:20 +08:00
  • 67149130be Add GPU utilization A/B test and fix cache-aware proxy bugs Gahow Wang 2026-05-21 22:13:38 +08:00
  • 05592e6adc Agentic workload PD separation analysis with trace-driven benchmarks Gahow Wang 2026-05-21 21:21:57 +08:00